Multi-head attention mechanism: “queries”, “keys”, and “values,” over and over again

This is the third article of my article series named “Instructions on Transformer for people outside NLP field, but with examples of NLP.”

In the last article, I explained how attention mechanism works in simple seq2seq models with RNNs, and it basically calculates correspondences of the hidden state at every time step, with all the outputs of the encoder. However I would say the attention mechanisms of RNN seq2seq models use only one standard for comparing them. Using only one standard is not enough for understanding languages, especially when you learn a foreign language. You would sometimes find it difficult to explain how to translate a word in your language to another language. Even if a pair of languages are very similar to each other, translating them cannot be simple switching of vocabulary. Usually a single token in one language is related to several tokens in the other language, and vice versa. How they correspond to each other depends on several criteria, for example “what”, “who”, “when”, “where”, “why”, and “how”. It is easy to imagine that you should compare tokens with several criteria.

Transformer model was first introduced in the original paper named “Attention Is All You Need,” and from the title you can easily see that attention mechanism plays important roles in this model. When you learn about Transformer model, you will see the figure below, which is used in the original paper on Transformer.  This is the simplified overall structure of one layer of Transformer model, and you stack this layer N times. In one layer of Transformer, there are three multi-head attention, which are displayed as boxes in orange. These are the very parts which compare the tokens on several standards. I made the head article of this article series inspired by this multi-head attention mechanism.

The figure below is also from the original paper on Transfromer. If you can understand how multi-head attention mechanism works with the explanations in the paper, and if you have no troubles understanding the codes in the official Tensorflow tutorial, I have to say this article is not for you. However I bet that is not true of majority of people, and at least I need one article to clearly explain how multi-head attention works. Please keep it in mind that this article covers only the architectures of the two figures below. However multi-head attention mechanisms are crucial components of Transformer model, and throughout this article, you would not only see how they work but also get a little control over it at an implementation level.

1 Multi-head attention mechanism

When you learn Transformer model, I recommend you first to pay attention to multi-head attention. And when you learn multi-head attentions, before seeing what scaled dot-product attention is, you should understand the whole structure of multi-head attention, which is at the right side of the figure above. In order to calculate attentions with a “query”, as I said in the last article, “you compare the ‘query’ with the ‘keys’ and get scores/weights for the ‘values.’ Each score/weight is in short the relevance between the ‘query’ and each ‘key’. And you reweight the ‘values’ with the scores/weights, and take the summation of the reweighted ‘values’.” Sooner or later, you will notice I would be just repeating these phrases over and over again throughout this article, in several ways.

*Even if you are not sure what “reweighting” means in this context, please keep reading. I think you would little by little see what it means especially in the next section.

The overall process of calculating multi-head attention, displayed in the figure above, is as follows (Please just keep reading. Please do not think too much.): first you split the V: “values”, K: “keys”, and Q: “queries”, and second you transform those divided “values”, “keys”, and “queries” with densely connected layers (“Linear” in the figure). Next you calculate attention weights and reweight the “values” and take the summation of the reiweighted “values”, and you concatenate the resulting summations. At the end you pass the concatenated “values” through another densely connected layers. The mechanism of scaled dot-product attention is just a matter of how to concretely calculate those attentions and reweight the “values”.

*In the last article I briefly mentioned that “keys” and “queries” can be in the same language. They can even be the same sentence in the same language, and in this case the resulting attentions are called self-attentions, which we are mainly going to see. I think most people calculate “self-attentions” unconsciously when they speak. You constantly care about what “she”, “it” , “the”, or “that” refers to in you own sentence, and we can say self-attention is how these everyday processes is implemented.

Let’s see the whole process of calculating multi-head attention at a little abstract level. From now on, we consider an example of calculating multi-head self-attentions, where the input is a sentence “Anthony Hopkins admired Michael Bay as a great director.” In this example, the number of tokens is 9, and each token is encoded as a 512-dimensional embedding vector. And the number of heads is 8. In this case, as you can see in the figure below, the input sentence “Anthony Hopkins admired Michael Bay as a great director.” is implemented as a 9\times 512 matrix. You first split each token into 512/8=64 dimensional, 8 vectors in total, as I colored in the figure below. In other words, the input matrix is divided into 8 colored chunks, which are all 9\times 64 matrices, but each colored matrix expresses the same sentence. And you calculate self-attentions of the input sentence independently in the 8 heads, and you reweight the “values” according to the attentions/weights. After this, you stack the sum of the reweighted “values”  in each colored head, and you concatenate the stacked tokens of each colored head. The size of each colored chunk does not change even after reweighting the tokens. According to Ashish Vaswani, who invented Transformer model, each head compare “queries” and “keys” on each standard. If the a Transformer model has 4 layers with 8-head multi-head attention , at least its encoder has 4\times 8 = 32 heads, so the encoder learn the relations of tokens of the input on 32 different standards.

I think you now have rough insight into how you calculate multi-head attentions. In the next section I am going to explain the process of reweighting the tokens, that is, I am finally going to explain what those colorful lines in the head image of this article series are.

*Each head is randomly initialized, so they learn to compare tokens with different criteria. The standards might be straightforward like “what” or “who”, or maybe much more complicated. In attention mechanisms in deep learning, you do not need feature engineering for setting such standards.

2 Calculating attentions and reweighting “values”

If you have read the last article or if you understand attention mechanism to some extent, you should already know that attention mechanism calculates attentions, or relevance between “queries” and “keys.” In the last article, I showed the idea of weights as a histogram, and in that case the “query” was the hidden state of the decoder at every time step, whereas the “keys” were the outputs of the encoder. In this section, I am going to explain attention mechanism in a more abstract way, and we consider comparing more general “tokens”, rather than concrete outputs of certain networks. In this section each [ \cdots ] denotes a token, which is usually an embedding vector in practice.

Please remember this mantra of attention mechanism: “you compare the ‘query’ with the ‘keys’ and get scores/weights for the ‘values.’ Each score/weight is in short the relevance between the ‘query’ and each ‘key’. And you reweight the ‘values’ with the scores/weights, and take the summation of the reweighted ‘values’.” The figure below shows an overview of a case where “Michael” is a query. In this case you compare the query with the “keys”, that is, the input sentence “Anthony Hopkins admired Michael Bay as a great director.” and you get the histogram of attentions/weights. Importantly the sum of the weights 1. With the attentions you have just calculated, you can reweight the “values,” which also denote the same input sentence. After that you can finally take a summation of the reweighted values. And you use this summation.

*I have been repeating the phrase “reweighting ‘values’  with attentions,”  but you in practice calculate the sum of those reweighted “values.”

Assume that compared to the “query”  token “Michael”, the weights of the “key” tokens “Anthony”, “Hopkins”, “admired”, “Michael”, “Bay”, “as”, “a”, “great”, and “director.” are respectively 0.06, 0.09, 0.05, 0.25, 0.18, 0.06, 0.09, 0.06, 0.15. In this case the sum of the reweighted token is 0.06″Anthony” + 0.09″Hopkins” + 0.05″admired” + 0.25″Michael” + 0.18″Bay” + 0.06″as” + 0.09″a” + 0.06″great” 0.15″director.”, and this sum is the what wee actually use.

*Of course the tokens are embedding vectors in practice. You calculate the reweighted vector in actual implementation.

You repeat this process for all the “queries.”  As you can see in the figure below, you get summations of 9 pairs of reweighted “values” because you use every token of the input sentence “Anthony Hopkins admired Michael Bay as a great director.” as a “query.” You stack the sum of reweighted “values” like the matrix in purple in the figure below, and this is the output of a one head multi-head attention.

3 Scaled-dot product

This section is a only a matter of linear algebra. Maybe this is not even so sophisticated as linear algebra. You just have to do lots of Excel-like operations. A tutorial on Transformer by Jay Alammar is also a very nice study material to understand this topic with simpler examples. I tried my best so that you can clearly understand multi-head attention at a more mathematical level, and all you need to know in order to read this section is how to calculate products of matrices or vectors, which you would see in the first some pages of textbooks on linear algebra.

We have seen that in order to calculate multi-head attentions, we prepare 8 pairs of “queries”, “keys” , and “values”, which I showed in 8 different colors in the figure in the first section. We calculate attentions and reweight “values” independently in 8 different heads, and in each head the reweighted “values” are calculated with this very simple formula of scaled dot-product: Attention(\boldsymbol{Q}, \boldsymbol{K}, \boldsymbol{V}) =softmax(\frac{\boldsymbol{Q} \boldsymbol{K} ^T}{\sqrt{d}_k})\boldsymbol{V}. Let’s take an example of calculating a scaled dot-product in the blue head.

At the left side of the figure below is a figure from the original paper on Transformer, which explains one-head of multi-head attention. If you have read through this article so far, the figure at the right side would be more straightforward to understand. You divide the input sentence into 8 chunks of matrices, and you independently put those chunks into eight head. In one head, you convert the input matrix by three different fully connected layers, which is “Linear” in the figure below, and prepare three matrices Q, K, V, which are “queries”, “keys”, and “values” respectively.

*Whichever color attention heads are in, the processes are all the same.

*You divide \frac{\boldsymbol{Q} \boldsymbol{K}} ^T by \sqrt{d}_k in the formula. According to the original paper, it is known that re-scaling \frac{\boldsymbol{Q} \boldsymbol{K}} ^T by \sqrt{d}_k is found to be effective. I am not going to discuss why in this article.

As you can see in the figure below, calculating Attention(\boldsymbol{Q}, \boldsymbol{K}, \boldsymbol{V}) is virtually just multiplying three matrices with the same size (Only K is transposed though). The resulting 9\times 64 matrix is the output of the head.

softmax(\frac{\boldsymbol{Q} \boldsymbol{K} ^T}{\sqrt{d}_k}) is calculated like in the figure below. The softmax function regularize each row of the re-scaled product \frac{\boldsymbol{Q} \boldsymbol{K} ^T}{\sqrt{d}_k}, and the resulting 9\times 9 matrix is a kind a heat map of self-attentions.

The process of comparing one “query” with “keys” is done with simple multiplication of a vector and a matrix, as you can see in the figure below. You can get a histogram of attentions for each query, and the resulting 9 dimensional vector is a list of attentions/weights, which is a list of blue circles in the figure below. That means, in Transformer model, you can compare a “query” and a “key” only by calculating an inner product. After re-scaling the vectors by dividing them with \sqrt{d_k} and regularizing them with a softmax function, you stack those vectors, and the stacked vectors is the heat map of attentions.

You can reweight “values” with the heat map of self-attentions, with simple multiplication. It would be more straightforward if you consider a transposed scaled dot-product \boldsymbol{V}^T \cdot softmax(\frac{\boldsymbol{Q} \boldsymbol{K} ^T}{\sqrt{d}_k})^T. This also should be easy to understand if you know basics of linear algebra.

One column of the resulting matrix (\boldsymbol{V}^T \cdot softmax(\frac{\boldsymbol{Q} \boldsymbol{K} ^T}{\sqrt{d}_k})^T) can be calculated with a simple multiplication of a matrix and a vector, as you can see in the figure below. This corresponds to the process or “taking a summation of reweighted ‘values’,” which I have been repeating. And I would like you to remember that you got those weights (blue) circles by comparing a “query” with “keys.”

Again and again, let’s repeat the mantra of attention mechanism together: “you compare the ‘query’ with the ‘keys’ and get scores/weights for the ‘values.’ Each score/weight is in short the relevance between the ‘query’ and each ‘key’. And you reweight the ‘values’ with the scores/weights, and take the summation of the reweighted ‘values’.” If you have been patient enough to follow my explanations, I bet you have got a clear view on how multi-head attention mechanism works.

We have been seeing the case of the blue head, but you can do exactly the same procedures in every head, at the same time, and this is what enables parallelization of multi-head attention mechanism. You concatenate the outputs of all the heads, and you put the concatenated matrix through a fully connected layers.

If you are reading this article from the beginning, I think this section is also showing the same idea which I have repeated, and I bet more or less you no have clearer views on how multi-head attention mechanism works. In the next section we are going to see how this is implemented.

4 Tensorflow implementation of multi-head attention

Let’s see how multi-head attention is implemented in the Tensorflow official tutorial. If you have read through this article so far, this should not be so difficult. I also added codes for displaying heat maps of self attentions. With the codes in this Github page, you can display self-attention heat maps for any input sentences in English.

The multi-head attention mechanism is implemented as below. If you understand Python codes and Tensorflow to some extent, I think this part is relatively easy.  The multi-head attention part is implemented as a class because you need to train weights of some fully connected layers. Whereas, scaled dot-product is just a function.

*I am going to explain the create_padding_mask() and create_look_ahead_mask() functions in upcoming articles. You do not need them this time.

Let’s see a case of using multi-head attention mechanism on a (1, 9, 512) sized input tensor, just as we have been considering in throughout this article. The first axis of (1, 9, 512) corresponds to the batch size, so this tensor is virtually a (9, 512) sized tensor, and this means the input is composed of 9 512-dimensional vectors. In the results below, you can see how the shape of input tensor changes after each procedure of calculating multi-head attention. Also you can see that the output of the multi-head attention is the same as the input, and you get a 9\times 9 matrix of attention heat maps of each attention head.

I guess the most complicated part of this implementation above is the split_head() function, especially if you do not understand tensor arithmetic. This part corresponds to splitting the input tensor to 8 different colored matrices as in one of the figures above. If you cannot understand what is going on in the function, I recommend you to prepare a sample tensor as below.

This is just a simple (1, 9, 512) sized tensor with sequential integer elements. The first row (1, 2, …., 512) corresponds to the first input token, and (4097, 4098, … , 4608) to the last one. You should try converting this sample tensor to see how multi-head attention is implemented. For example you can try the operations below.

These operations correspond to splitting the input into 8 heads, whose sizes are all (9, 64). And the second axis of the resulting (1, 8, 9, 64) tensor corresponds to the index of the heads. Thus sample_sentence[0][0] corresponds to the first head, the blue 9\times 64 matrix. Some Tensorflow functions enable linear calculations in each attention head, independently as in the codes below.

Very importantly, we have been only considering the cases of calculating self attentions, where all “queries”, “keys”, and “values” come from the same sentence in the same language. However, as I showed in the last article, usually “queries” are in a different language from “keys” and “values” in translation tasks, and “keys” and “values” are in the same language. And as you can imagine, usualy “queries” have different number of tokens from “keys” or “values.” You also need to understand this case, which is not calculating self-attentions. If you have followed this article so far, this case is not that hard to you. Let’s briefly see an example where the input sentence in the source language is composed 9 tokens, on the other hand the output is composed 12 tokens.

As I mentioned, one of the outputs of each multi-head attention class is 9\times 9 matrix of attention heat maps, which I displayed as a matrix composed of blue circles in the last section. The the implementation in the Tensorflow official tutorial, I have added codes to display actual heat maps of any input sentences in English.

*If you want to try displaying them by yourself, download or just copy and paste codes in this Github page. Please maker “datasets” directory in the same directory as the code. Please download “” from this page, and unzip it. After that please put “spa.txt” on the “datasets” directory. Also, please download the “checkpoints_en_es” folder from this link, and place the folder in the same directory as the file in the Github page. In the upcoming articles, you would need similar processes to run my codes.

After running codes in the Github page, you can display heat maps of self attentions. Let’s input the sentence “Anthony Hopkins admired Michael Bay as a great director.” You would get a heat maps like this.

In fact, my toy implementation cannot handle proper nouns such as “Anthony” or “Michael.” Then let’s consider a simple input sentence “He admired her as a great director.” In each layer, you respectively get 8 self-attention heat maps.

I think we can see some tendencies in those heat maps. The heat maps in the early layers, which are close to the input, are blurry. And the distributions of the heat maps come to concentrate more or less diagonally. At the end, presumably they learn to pay attention to the start and the end of sentences.

You have finally finished reading this article. Congratulations.

You should be proud of having been patient, and you passed the most tiresome part of learning Transformer model. You must be ready for making a toy English-German translator in the upcoming articles. Also I am sure you have understood that Michael Bay is a great director, no matter what people say.

*Hannibal Lecter, I mean Athony Hopkins, also wrote a letter to the staff of “Breaking Bad,” and he told them the tv show let him regain his passion. He is a kind of admiring around, and I am a little worried that he might be getting senile. He played a role of a father forgetting his daughter in his new film “The Father.” I must see it to check if that is really an acting, or not.


[1] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin, “Attention Is All You Need” (2017)

[2] “Transformer model for language understanding,” Tensorflow Core

[3] “Neural machine translation with attention,” Tensorflow Core

[4] Jay Alammar, “The Illustrated Transformer,”

[5] “Stanford CS224N: NLP with Deep Learning | Winter 2019 | Lecture 14 – Transformers and Self-Attention,” stanfordonline, (2019)

[6]Tsuboi Yuuta, Unno Yuuya, Suzuki Jun, “Machine Learning Professional Series: Natural Language Processing with Deep Learning,” (2017), pp. 91-94
坪井祐太、海野裕也、鈴木潤 著, 「機械学習プロフェッショナルシリーズ 深層学習による自然言語処理」, (2017), pp. 191-193

[7]”Stanford CS224N: NLP with Deep Learning | Winter 2019 | Lecture 8 – Translation, Seq2Seq, Attention”, stanfordonline, (2019)

[8]Rosemary Rossi, “Anthony Hopkins Compares ‘Genius’ Michael Bay to Spielberg, Scorsese,” yahoo! entertainment, (2017)

* I make study materials on machine learning, sponsored by DATANOMIQ. I do my best to make my content as straightforward but as precise as possible. I include all of my reference sources. If you notice any mistakes in my materials, including grammatical errors, please let me know (email: And if you have any advice for making my materials more understandable to learners, I would appreciate hearing it.

Data Mining Process flow – Easy Understanding

1 Overview

Development of computer processing power, network and automated software completely change and give new concept of each business. And data mining play the vital part to solve, finding the hidden patterns and relationship from large dataset with business by using sophisticated data analysis tools like methodology, method, process flow etc.

On this paper, proposed a process flow followed CRISP-DM methodology and has six steps where data understanding does not considered.

Phase of new process flow given below:-

Phase 1: Involved with collection, outliner treatment, imputation, transformation, scaling, and partition dataset in to two sub-frames (Training and Testing). Here as an example for outliner treatment, imputation, transformation, scaling consider accordingly Z score, mean, One hot encoding and Min Max Scaler.

Phase 2: On this Phase training and testing data balance with same balancing algorithm but separately. As an example here SMOTE (synthetic minority oversampling technique) is considered.

Phase 3: This phase involved with reduction, selection, aggregation, extraction. But here for an example considering same feature reduction algorithm (LDA -Linear Discriminant analysis) on training and testing data set separately.

Phase 4: On this Phase Training data set again partition into two more set (Training and Validation).

Phase 5: This Phase considering several base algorithms as a base model like CNN, RNN, Random forest, MLP, Regression, Ensemble method. This phase also involve to find out best hyper parameter and sub-algorithm for each base algorithm. As an example on this paper consider two class classification problems and also consider Random forest (Included CART – Classification and Regression Tree and GINI index impurity) and MLP classifier (Included (Relu, Sigmoid, binary cross entropy, Adam – Adaptive Moment Estimation) as base algorithms.

Phase 6: First, Prediction with validation data then evaluates with Test dataset which is fully unknown for these (Random forest, MLP classifier) two base algorithms. Then calculate the confusion matrix, ROC, AUC to find the best base algorithm.

New method from phase 1 to phase 4 followed CRISP-DM methodology steps such as data collection, data preparation then phase 5 followed modelling and phase 6 followed evaluation and implementation steps.

Structure of proposed process flow for two class problem combined with algorithm and sub-algorithm display on figure – 1.

These articles mainly focus to describe all algorithms which are going to implementation for better understanding.



Data Mining Process Flow

Figure 1 – Data Mining Process Flow

2 Phase 1: Outlier treatment, Transform, Scaling, Imputation

This phase involved with outlier treatment, imputation, scaling, and transform data.

2.1 Outliner treatment: – Z score

Outlier is a data point which lies far from all other data point in a data set. Outlier need to treat because it may bias the entire result. Outlier treatment with Z score is a common technique.  Z score is a standard score in statistics.  Z score provides information about data value is smaller or grater then mean that means how many standard deviations away from the mean value. Z score equation display below:

Z = \frac{(x - \mu)}{\sigma}

Here x = data point
σ = Standard deviation
μ = mean value

Equation- 1 Z-Score

In a normal distribution Z score represent 68% data lies on +/- 1, 95% data point lies on +/- 2, 99.7% data point lies on +/- 3 standard deviation.

2.2 Imputation data: – mean

Imputation is a way to handle missing data by replacing substituted value. There are many imputation technique represent like mean, median, mode, k-nearest neighbours. Mean imputation is the technique to replacing missing information with mean value. On the mean imputation first calculate the particular features mean value and then replace the missing value with mean value. The next equation displays the mean calculation:

\mu = \frac{(\sum x)}{n}

Here x = value of each point
n = number of values
μ = mean value

Equation- 2 Mean

2.3 Transform: – One hot encoding

Encoding is a pre-processing technique which represents data in such a way that computer can understand.  For understanding of machine learning algorithm categorical columns convert to numerical columns, this process called categorical encoding. There are multiple way to handle categorical variable but most widely used techniques are label encoding and one host encoding. On label encoding give a numeric (integer number) for each category. Suppose there are 3 categories of foods like apples, orange, banana. When label encoding is used then 3 categories will get a numerical value like apples = 1, banana = 2 and orange = 3. But there is very high probability that machine learning model can capture the relationship in between categories such as apple < banana < orange or calculate average across categories like 1 +3 = 4 / 2 = 2 that means model can understand average of apple and orange together is banana which is not acceptable because model correlation calculation is wrong. For solving this problem one hot encoding appear. The following table displays the label encoding is transformed into one hot encoding.

Label Encoding and One-Hot-Encoding

Table- 1 Encoding example

On hot encoding categorical value split into columns and each column contains 0 or 1 according to columns placement.

2.4 Scaling data: – Min Max Scaler

Feature scaling method is standardized or normalization the independent variable that means it is used to scale the data in a particular range like -1 to +1 or depending on algorithm. Generally normalization used where data distribution does not follow Gaussian distribution and standardization used where data distribution follow Gaussian distribution. On standardization techniques transform data values are cantered around the mean and unit is standard deviation. Formula for standardization given below:

Standardization X = \frac{(X - \mu)}{\sigma}

Equation-3 Equations for Standardization

X represent the feature value, µ represent mean of the feature value and σ represent standard deviation of the feature value. Standardized data value does not restrict to a particular range.

Normalization techniques shifted and rescaled data value range between 0 and 1. Normalization techniques also called Min-Max scaling. Formula for normalization given below:

Normalization X = \frac{(X - X_{min})}{X_{max} - X_{min}}

Equation – 4 Equations for Normalization

Above X, Xmin, Xmax are accordingly feature values, feature minimum value and feature maximum value. On above formula when X is minimums then numerator will be 0 (  is 0) or if X is maximums then the numerator is equal to the denominator (  is 1). But when X data value between minimum and maximum then  is between 0 and 1. If ranges value of data does not normalized then bigger range can influence the result.

3 Phase 2: – Balance Data


SMOTE (synthetic minority oversampling technique) is an oversampling technique where synthetic observations are created based on existing minority observations. This technique operates in feature space instead of data space. Under SMOTE each minority class observation calculates k nearest neighbours and randomly chose the neighbours depending on over-sampling requirements. Suppose there are 4 data point on minority class and 10 data point on majority class. For this imbalance data set, balance by increasing minority class with synthetic data point.   SMOTE creating synthetic data point but it is necessary to consider k nearest neighbours first. If k = 3 then SMOTE consider 3 nearest neighbours. Figure-2 display SMOTE with k = 3 and x = x1, x2, x3, x4 data point denote minority class. And all circles represent majority class.

SMOTE Example

Figure- 2 SMOTE example


4 Phase 3: – Feature Reduction

4.1 LDA

LDA stands for Linear Discriminant analysis supervised technique are commonly used for classification problem.  On this feature reduction account continuous independent variable and output categorical variable. It is multivariate analysis technique. LDA analyse by comparing mean of the variables.  Main goal of LDA is differentiate classes in low dimension space. LDA is similar to PCA (Principal component analysis) but in addition LDA maximize the separation between multiple classes. LDA is a dimensionality reduction technique where creating synthetic feature from linear combination of original data set then discard less important feature. LDA calculate class variance, it maximize between class variance and minimize within class variance. Table-2 display the process steps of LDA.

LDA Process

Table- 2 LDA process

5 Phase 5: – Base Model

Here we consider two base model ensemble random forest and MLP classifier.

5.1 Random Forest

Random forest is an ensemble (Bagging) method where group of weak learner (decision tree) come together to form a strong leaner. Random forest is a supervised algorithm which is used for regression and classification problem. Random forests create several decisions tree for predictions and provide solution by voting (classification) or mean (regression) value. Working process of Random forest given below (Table -3).

Random Forest

Table-3 Random Forest process

When training a Random forest root node contains a sample of bootstrap dataset and the feature is as same as original dataset. Suppose the dataset is D and contain d record and m number of columns. From the dataset D random forest first randomly select sample of rows (d) with replacement and sample of features (n) and give it to the decision tree. Suppose Random forest created several decision trees like T1, T2, T3, T4 . . . Tn. Then randomly selected dataset D = d + n is given to the decision tree T1, T2, T3, T4 . . . Tn where D < D, m > n and d > d.  After taking the dataset decision tree give the prediction for binary classification 1 or 0 then aggregating the decision and select the majority voted result. Figure-3 describes the structure of random forest process.

Random Forest Process

Figure- 3 Random Forest process

On Random forest base learner Decision Tree grows complete depth where bias (properly train on training dataset) is low and variance is high (when implementing test data give big error) called overfitting. On Random forest using multiple decision trees where each Decision tree is high variance but when combining all decision trees with the respect of majority vote then high variance converted into low variance because using row and feature sampling with replacement and taking the majority vote where decision is not depend on one decision tree.

CART (Classification and Regression Tree) is binary segmentation technique. CART is a Gini’s impurity index based classical algorithm to split a dataset and build a decision tree. By splitting a selected dataset CART created two child nodes repeatedly and builds a tree until the data no longer be split. There are three steps CART algorithm follow:

  1. Find best split for each features. For each feature in binary split make two groups of the ordered classes. That means possibility of split for k classes is k-1. Find which split is maximized and contain best splits (one for each feature) result.
  2. Find the best split for nodes. From step 1 find the best one split (from all features) which maximized the splitting criterion.
  3. Split the best node from step 2 and repeat from step 1 until fulfil the stopping criterion.


For splitting criteria CART use GINI index impurity algorithm to calculate the purity of split in a decision tree. Gini impurity randomly classified the labels with the same distribution in the dataset. A Gini impurity of 0 (lowest) is the best possible impurity and it is achieve when everything is in a same class. Gini index varies from 0 to 1. 0 indicate the purity of class where only one class exits or all element under a specific class. 1 indicates that elements are randomly distributed across various classes. And 0.5 indicate equal elements distributed over classes. Gini index (GI) described by mathematically that sum of squared of probabilities of each class (pi) deducted from one (Equation-5).

Gini Impurities

Equation – 5 Gini impurities

Here (Equation-5) pi represent the probability (probability of p+ or yes and probability of p- or no) of distinct class with classified element. Suppose randomly selected feature (a1) which has 8 yes and 4 no. After the split right had side (b1 on equation-6) has 4 yes and 4 no and left had side (b2 on equation – 7) has 4 yes and 0 no. here b2 is a pure split (leaf node) because only one class yes is present. By using the GI (Gini index) formula for b1 and b2:-

Equation- 6 & 7 – Gini Impurity b1 & Gini Impurity b2

Here for b1 value 0.5 indicates that equal element (yes and no) distribute over classes which is not pure split. And b2 value 0 indicates pure split. On GINI impurity indicates that when probability (yes or no) increases GINI value also increases. Here 0 indicate pure split and .5 indicate equal split that means worst situation. After calculating the GINI index for b1 and b2 now calculate the reduction of impurity for data point a1. Here total yes 8 (b1 and b2 on Equation – 8) and total no 4 (b1) so total data is 12 on a1. Below display the weighted GINI index for feature a1:

Total data point on b1 with Gini index (m) = 8/12 * 0.5 = 0.3333

Total data point on b2 with Gini index (n) = 4/12 * 0 = 0

Weighted Gini index for feature a1 = m + n = 0.3333

Equation- 8 Gini Impurity b1 & b2

After computing the weighted Gini value for every feature on a dataset taking the highest value feature as first node and split accordingly in a decision tree. Gini is less costly to compute.

5.2 Multilayer Perceptron Classifier (MLP Classifier)

Multilayer perceptron classifier is a feedforward neural network utilizes supervised learning technique (backpropagation) for training. MLP Classifier combines with multiple perceptron (hidden) layers. For feedforward taking input send combining with weight bias and then activation function from one hidden layer output goes to other hidden and this process continuing until reached the output. Then output calculates the error with error algorithm. These errors send back with backpropagation for weight adjustment by decreasing the total error and process is repeated, this process is call epoch. Number of epoch is determined with the hyper-parameter and reduction rate of total error.

5.2.1 Back-Propagation

Backpropagation is supervised learning algorithm that is used to train neural network. A neural network consists of input layer, hidden layer and output layer and each layer consists of neuron. So a neural network is a circuit of neurons. Backpropagation is a method to train multilayer neural network the updating of the weights of neural network and is done in such a way so that the error observed can be reduced here, error is only observed in the output layer and that error is back propagated to the previous layers and previous layer is proportionally updated weight. Backpropagation maintain chain rule to update weight. Mainly three steps on backpropagation are (Table-4):

Step Process
Step 1 Forward Pass
Step 2 Backward Pass
Step 3 Sum of all values and calculate updated weight value with Chain – rules.

Table-4 Back-Propagation process

5.2.2 Forward pass/ Forward propagation

Forward propagation is the process where input layer send the input value with randomly selected weight and bias to connected neuron and inside neuron selected activation function combine them and forward to other connected neuron layer after layer then give an output with the help of output layer. Below (Figure-4) display the forward propagation.

Foreward Pass

Figure-4 Forward passes

Input layer take the input of X (X1, X2) combine with randomly selected weight for each connection and with fixed bias (different hidden layer has different bias) send it to first hidden layer where first multiply the input with corresponding weight and added all input with single bias then selected activation function (may different form other layer) combine all input and give output according to function and this process is going on until reach in output layer. Output layer give the output like Y (Y1, Y2) (here output is binary classification as an example) according to selected activation function.

5.2.3 Backward Pass

After calculating error (difference between Forward pass output and actual output) backward pass try to minimize the error with optimisation function by sending backward with proportionally distribution and maintain a chain rule. Backward pass distribution the error in such a way where weighted value is taking under consideration. Below (Figure-5) diagram display the Backward pass process.

Backward Pass

Figure-5 Backward passes

Backpropagation push back the error which is calculated with error function or loss function for update proportional distribution with the help of optimisation algorithm. Division of Optimisation algorithm given below on Figure – 6

Optimisation Algorithms

Figure -6 Division of Optimisation algorithms

Gradient decent calculate gradient and update value by increases or decreases opposite direction of gradients unit and try to find the minimal value. Gradient decent update just one time for whole dataset but stochastic gradient decent update on each training sample and it is faster than normal gradient decent. Gradient decent can be improve by tuning parameter like learning rate (0 to 1 mostly use 0.5). Adagrad use time step based parameter to compute learning rate for every parameter. Adam is Adaptive Moment Estimation. It calculates different parameter with different learning rate. It is faster and performance rate is higher than other optimization algorithm. On the other way Adam algorithm is squares the calculated exponential weighted moving average of gradient.

5.2.4 Chain – rules

Backpropagation maintain chain-rules to update weighted value. On chain-rules backpropagation find the derivative of error respect to any weight. Suppose E is output error. w is weight for input a and bias b and ac neuron output respect of activation function and summation of bias with weighted input (w*a) input to neuron is net. So partial derivative for error respect to weight is ∂E / ∂w display the process on figure-7.

Figure- 7 Partial derivative for error respect to weight

On the chain rules for backward pass to find (error respect to weight) ∂E / ∂w = ∂E / ∂ac * ∂ac / ∂net * ∂net / ∂w. here find to error respect to weight are error respect to output of activation function multiply by activation function output respect to input in a neuron multiply by input in a neuron respect to weight.

5.2.5 Activation function

Activation function is a function which takes the decision about neuron to activate or deactivate. If the activate function activate the neuron then it will give an output on the basis of input. Input in a activation function is sum of input multiply with corresponding weight and adding the layered bias.  The main function of a activate function is non-linearity output of a neuron.

Activation Function

Figure-8 Activation function

Figure – 8 display a neuron in a hidden layer. Here several input (1, 2, 3) with corresponding weight (w1, w2, w3) putting in a neuron input layer where layer bias add with summation of multiplication with input and weight. Equation-9 display the output of an activate function.

Output from activate function y = Activate function (Ʃ (weight * input) + bias)

y = f (Ʃ (w*x) +b)

Equation- 9 Activate function

There are many activation functions like linear function for regression problem, sigmoid function for binary classification problem where result either 0 or 1, Tanh function which is based on sigmoid function but mathematically shifted version and values line -1 to 1. RELU function is Rectified linear unit. RELU is less expensive to compute.

5.2.6 Sigmoid

Sigmoid is a squashing activate function where output range between 0 and 1. Sigmoidal name comes from Greek letter sigma which looks like letter S when graphed. Sigmoid function is a logistic type function, it mainly use in output layer in neural network. Sigmoid is non-linear, fixed output range (between 0 and 1), monotonic (never decrees or never increases) and continuously differentiated function. Sigmoid function is good at classification and output from sigmoid is nonlinear. But Sigmoid has a vanishing gradient problem because output variable is very less to change in input variable. Figure- 9 displays the output of a Sigmoid and derivative of Sigmoid. Here x is any number (positive or negative). On sigmoid function 1 is divided by exponential negative input with adding 1.


Figure – 9 Sigmoid Functions RELU

RELU stands for Rectified Linear Units it is simple, less expensive in computation and rectifies the gradient vanishing problem. RELU is nonlinear activation function. It gives output either positive (infinity) or 0. RELU has a dying problem because if neurons stop for responding to variation because of gradient is 0 or nothing has to change. Figure- 10 displays the output of an RELU and derivative of RELU. Here x is any positive input and if x is grater then 0 give the output as x or give output 0. RELU function gives the output maximum value of input, here max (0, x).

Relu Activation Function

Figure – 10 RELU Function Cost / loss function (Binary Cross-Entropy)

Cost or loss function compare the predictive value (model outcome) with actual value and give a quantitative value which give the indication about how much good or bad the prediction is.

Cost Function

Figure- 11 Cost function work process

Figure-11 x1 and x2 are input in a activate function f(x) and output y1_out which is sum of weighted input added with bias going through activate function. After model output activate function compare the output with actual output and give a quantitative value which indicate how good or bad the prediction is.

There are many type of loss function but choosing of optimal loss function depends on the problem going to be solved such as regression or classification. For binary classification problem binary cross entropy is used to calculate cost. Equation-10 displays the binary cross entropy where y is actual binary value and yp predictive outcome range 0 and 1. And i is scalar vale range between 1 to model output size (N).

Binary Crossentropy

Equation-10 displays the binary cross entropy

6 Phase 6: – Evaluation

6.1 Confusion matrix

In a classification confusion matrix describe the performance of actual value against predictive value. Confusion Matrix does the performance measurement. So confusion matrix classifies and display predicted and actual value (Visa, S., Ramsay 2011).

Confusion Matrix

Table- 5 Confusion Matrix

Confusion Matrix (Table-5) combines with True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN). True Positive is prediction positive and true. True Negative is prediction negative and that is true. False positive is prediction positive and it’s false. False negative is prediction negative and that is false. False positive is known as Type1 error and false negative is known as Type 2 error. Confusion matrix can able to calculate several list of rates which are given below on Table- 6.

Here    N = Total number of observation, TP = True Positive, TN = True Negative

FP = False Positive, FN = False Negative, Total Actual No (AN) = TN + FP,

Total Predictive Yes (PY) = FP + TP. Total Actual Yes (AY) = FN + TP



Description Mathematical Description
Accuracy Classifier, overall how often correctly identified  (TP+TN) / N
Misclassification Rate Classifier, overall how often wrongly identified (FP + FN) / N
True Positive Rate

(Sensitivity / Recall)

Classifier, how often predict correctly yes when it is actually yes.  TP / AY
False Positive Rate Classifier, how often predict wrongly yes when it is actually no.  FP / AN
True Negative Rate


Classifier, how often predict correctly no when it is actually no.  TN / AN
Precision Classifier how often predict yes when it is correct.  TP / PY
Prevalence Yes conditions how often occur in a sample. AY / N

Table – 6 Confusion matrixes Calculation

From confusion matrix F1 score can be calculated because F1 score related to precision and recall. Higher F1 score is better. If precision or recall any one goes down F1 score also go down.

F1 = \frac{2 * Precision * Recall}{Precision + Recall}

4.6.2 ROC (Receiver Operating Characteristic) curve

In statistics ROC is represent in a graph with plotting a curve which describe a binary classifiers performance as its differentiation threshold is varied. ROC (Equation-11) curve created true positive rate (TPR) against false positive rate (FPR). True positive rate also called as Sensitivity and False positive rate also known as Probability of false alarm. False positive rate also called as a probability of false alarm and it is calculated as 1 – Specificity.

True Positive Rate = \frac{True Positive}{True Positive + False Negative} = Recall or Sensivity

False Positive Rate = \frac{True Negative}{True Negative + False Positive} = 1 - Specificity

Equation- 11 ROC

So ROC (Receiver Operation Characteristic) curve allows visual representation between sensitivity and specificity associated with different values of the test result (Grzybowski, M. and Younger, J.G., 1997)

On ROC curve each point has different Threshold level. Below (Figure – 12) display the ROC curve. Higher the area curve covers is better that means high sensitivity and high specificity represent more accuracy. ROC curve also represent that if classifier predict more often true than it has more true positive and also more false positive. If classifier predict true less often then fewer false positive and also fewer true positive.

ROC Curve

ROC Curve

Figure – 12 ROC curve description

4.6.3 AUC (Area under Curve)

Area under curve (AUC) is the area surrounded by the ROC curve and AUC also represent the degree of separability that means how good the model to distinguished between classes. Higher the AUC value represents better the model performance to separate classes. AUC = 1 for perfect classifier, AUC = 0 represent worst classifier, and AUC = 0.5 means has no class separation capacity. Suppose AUC value is 0.6 that means 60% chance that model can classify positive and negative class.

Figure- 13 to Figure – 16 displays an example of AUC where green distribution curve for positive class and blue distribution curve for negative class. Here threshold or cut-off value is 0.5 and range between ‘0’ to ‘1’. True negative = TN, True Positive = TP, False Negative = FN, False Positive = FP, True positive rate = TPR (range 0 to 1), False positive rate = FPR (range 0 to 1).

On Figure – 13 left distribution curve where two class curves does not overlap that means both class are perfectly distinguished. So this is ideal position and AUC value is 1.  On the left side ROC also display that TPR for positive class is 100% occupied.

ROC distributions (perfectly distinguished

ROC distributions (perfectly distinguished

Figure – 14 two class overlap each other and raise false positive (Type 1), false negative (Type 2) errors. Here error could be minimize or maximize according to threshold. Suppose here AUC = 0.6, that means chance of a model to distinguish two classes is 60%. On ROC curve also display the curve occupied for positive class is 60%.

ROC distributions (class partly overlap distinguished)

ROC distributions (class partly overlap distinguished)

Figure- 15 displayed that positive and negative overlap each other. Here AUC value is 0.5 or near to 0.5. On this position classifier model does not able distinguish positive and negative classes. On left side ROC curve become straight that means TPR and FPR are equal.

ROC distributions (class fully overlap distinguished)

ROC distributions (class fully overlap distinguished)

Figure- 16 positive and negative class swap position and on this position AUC = 0. That means classified model predict positive as a negative and negative as a positive. On the left ROC curve display that curve on FPR side fully fitted.

ROC distributions (class swap position distinguished)

ROC distributions (class swap position distinguished)

7 Summaries

This paper describes a data mining process flow and related model and its algorithm with textual representation. One hot encoding create dummy variable for class features and min-max scaling scale the data in a single format. Balancing by SMOTE data where Euclidian distance calculates the distance in-between nearest neighbour to produce synthetic data under minority class. LDA reduce the distance inside class and maximise distance in-between class and for two class problem give a single dimension features which is less costly to calculate accuracy by base algorithm (random forest and MLP classifier).  Confusion matrix gives the accuracy, precision, sensitivity, specificity which is help to take a decision about base algorithm. AUC and ROC curve also represent true positive rate against false positive rate which indicate base algorithm performance.

Base algorithm Random forest using CART with GINI impurity for feature selection to spread the tree. Here CART is selected because of less costly to run. Random forest algorithm is using bootstrap dataset to grow trees, and aggregation using majority vote to select accuracy.

MLP classifier is a neural network algorithm using backpropagation chain-rule to reducing error. Here inside layers using RLU activation function. Output layers using Sigmoid activation function and binary cross entropy loss function calculate the loss which is back propagate with Adam optimizer to optimize weight and reduce loss.


  1. Visa, S., Ramsay, B., Ralescu, A.L. and Van Der Knaap, E., 2011. Confusion Matrix-based Feature Selection. MAICS, 710, pp.120-127.
  2. Grzybowski, M. and Younger, J.G., 1997. Statistical methodology: III. Receiver operating characteristic (ROC) curves. Academic Emergency Medicine, 4(8), pp.818-826.

On the difficulty of language: prerequisites for NLP with deep learning

This is the first article of my article series “Instructions on Transformer for people outside NLP field, but with examples of NLP.”

1 Preface

This section is virtually just my essay on language. You can skip this if you want to get down on more technical topic.

As I do not study in natural language processing (NLP) field, I would not be able to provide that deep insight into this fast changing deep leaning field throughout my article series. However at least I do understand language is a difficult and profound field, not only in engineering but also in many other study fields. Some people might be feeling that technologies are eliminating languages, or one’s motivations to understand other cultures. First of all, I would like you to keep it in mind that I am not a geek who is trying to turn this multilingual world into a homogeneous one and rebuild Tower of Babel, with deep learning. I would say I am more keen on social or anthropological sides of language.

I think you would think more about languages if you have mastered at least one foreign language. As my mother tongue is Japanese, which is totally different from many other Western languages in terms of characters and ambiguity, I understand translating is not what learning a language is all about. Each language has unique characteristics, and I believe they more or less influence one’s personalities. For example, many Western languages make the verb, I mean the conclusion, of sentences clear in the beginning part of the sentences. That is also true of Chinese, I heard. However in Japanese, the conclusion comes at the end, so that is likely to give an impression that Japanese people are being obscure or indecisive. Also, Japanese sentences usually omit their subjects. In German as well, the conclusion of a sentences tend to come at the end, but I am almost 100% sure that no Japanese people would feel German people make things unclear. I think that comes from the structures of German language, which tends to make the number, verb, relations of words crystal clear.

Let’s take an example to see how obscure Japanese is. A Japanese sentence 「頭が赤い魚を食べる猫」can be interpreted in five ways, depending on where you put emphases on.

Common sense tells you that the sentence is likely to mean the first two cases, but I am sure they can mean those five possibilities. There might be similarly obscure sentences in other languages, but I bet few languages can be as obscure as Japanese. Also as you can see from the last two sentences, you can omit subjects in Japanese. This rule is nothing exceptional. Japanese people usually don’t use subjects in normal conversations. And when you read classical Japanese, which Japanese high school students have to do just like Western students learn some of classical Latin, the writings omit subjects much more frequently.

*However interestingly we have rich vocabulary of subjects. The subject “I” can be translated to 「私」、「僕」、「俺」、「自分」、「うち」etc, depending on your personality, who you are talking to, and the time when it is written in.

I believe one can see the world only in the framework of their language, and it seems one’s personality changes depending on the language they use. I am not sure whether the language originally determines how they think, or how they think forms the language. But at least I would like you to keep it in mind that if you translate a conversation, for example a random conversation at a bar in Berlin, into Japanese, that would linguistically sound Japanese, but not anthropologically. Imagine that such kind of random conversation in Berlin or something is like playing a catch, I mean throwing a ball named “your opinion.” On the other hand,  normal conversations of Japanese people are in stead more of, I would say,  “resonance” of several tuning forks. They do their bests to show that they are listening to each other, by excessively nodding or just repeating “Really?”, but usually it seems hardly any constructive dialogues have been made.

*I sometimes feel you do not even need deep learning to simulate most of such Japanese conversations. Several-line Python codes would be enough.

My point is, this article series is mainly going to cover only a few techniques of NLP in deep learning field: sequence to sequence model (seq2seq model) , and especially Transformer. They are, at least for now, just mathematical models and mappings of a small part of this profound field of language (as far as I can cover in this article series). But still, examples of language would definitely help you understand Transformer model in the long run.

2 Tokens and word embedding

*Throughout my article series, “words” just means the normal words you use in daily life. “Tokens” means more general unit of NLP tasks. For example the word “Transformer” might be denoted as a single token “Transformer,” or maybe as a combination of two tokens “Trans” and “former.”

One challenging part of handling language data is its encodings. If you started learning programming in a language other than English, you would have encountered some troubles of using keyboards with different arrangements or with characters. Some comments on your codes in your native languages are sometimes not readable on some software. You can easily get away with that by using only English, but when it comes to NLP you have to deal with this difficulty seriously. How to encode characters in each language should be a first obstacle of NLP. In this article we are going to rely on a library named BPEmb, which provides word embedding in various languages, and you do not have to care so much about encodings in languages all over the world with this library.

In the first section, you might have noticed that Japanese sentence is not separated with spaces like Western languages. This is also true of Chinese language, and that means we need additional tasks of separating those sentences at least into proper chunks of words. This is not only a matter of engineering, but also of some linguistic fields. Also I think many people are not so conscious of how sentences in their native languages are grammatically separated.

The next point is, unlike other scientific data, such as temperature, velocity, voltage, or air pressure, language itself is not measured as numerical data. Thus in order to process language, including English, you first have to map language to certain numerical data, and after some processes you need to conversely map the output numerical data into language data. This section is going to be mainly about one-hot encoding and word embedding, the ways to convert word/token into numerical data. You might already have heard about this

You might have learnt about word embedding to some extent, but I hope you could get richer insight into this topic through this article.

2.1 One-hot encoding

One-hot encoding would be the most straightforward way to encode words/tokens. Assume that you have a dictionary whose size is |\mathcal{V}|, and it includes words from “a”, “ablation”, “actually” to “zombie”, “?”, “!”

In a mathematical manner, in order to choose a word out of those |\mathcal{V}| words, all you need is a |\mathcal{V}| dimensional vector, one of whose elements is 1, and the others are 0. When you want to choose the No. i word, which is “indeed” in the example below, its corresponding one-hot vector is \boldsymbol{v} = (0, \dots, 1, \dots, 0 ), where only the No. i element is 1. One-hot encoding is also easy to understand, and that’s all. It is easy to imagine that people have already come up with more complicated and better way to encoder words. And one major way to do that is word embedding.

2.2 Word embedding

Source: Francois Chollet, Deep Learning with Python,(2018), Manning

Actually word embedding is related to one-hot encoding, and if you understand how to train a simple neural network, for example densely connected layers, you would understand word embedding easily. The key idea of word embedding is denoting each token with a D dimensional vector, whose dimension is fewer than the vocabulary size |\mathcal{V}|. The elements of the resulting word embedding vector are real values, I mean not only 0 or 1. Obviously you can encode much richer variety of tokens with such vectors. The figure at the left side is from “Deep Learning with Python” by François Chollet, and I think this is an almost perfect and simple explanation of the comparison of one-hot encoding and word embedding. But the problem is how to get such convenient vectors. The answer is very simple: you have only to train a network whose inputs are one-hot vector of the vocabulary.

The figure below is a simplified model of word embedding of a certain word. When the word is input into a neural network, only the corresponding element of the one-hot vector is 1, and that virtually means the very first input layer is composed of one neuron whose value is 1. And the only one neuron propagates to the next D dimensional embedding layer. These weights are the very values which most other study materials call “an embedding vector.”

When you input each word into a certain network, for example RNN or Transformer, you map the input one-hot vector into the embedding layer/vector. The examples in the figure are how inputs are made when the input sentences are “You’ve got the touch” and “You’ve got the power.”   Assume that you have a dictionary of one-hot encoding, whose vocabulary is {“the”, “You’ve”, “Walberg”, “touch”, “power”, “Nights”, “got”, “Mark”, “Boogie”}, and the dimension of word embeding is 6. In this case |\mathcal{V}| = 9, D=6. When the inputs are “You’ve got the touch” or “You’ve got the power” , you put the one-hot vector corresponding to “You’ve”, “got”, “the”, “touch” or “You’ve”, “got”, “the”, “power” sequentially every time step t.

In order to get word embedding of certain vocabulary, you just need to train the network. We know that the words “actually” and “indeed” are used in similar ways in writings. Thus when we propagate those words into the embedding layer, we can expect that those embedding layers are similar. This is how we can mathematically get effective word embedding of certain vocabulary.

More interestingly, if word embedding is properly trained, you can mathematically “calculate” words. For example, \boldsymbol{v}_{king} - \boldsymbol{v}_{man} + \boldsymbol{v}_{woman} \approx \boldsymbol{v}_{queen}, \boldsymbol{v}_{Japan} - \boldsymbol{v}_{Tokyo} + \boldsymbol{v}_{Vietnam} \approx \boldsymbol{v}_{Hanoi}.

*I have tried to demonstrate this type of calculation on several word embedding, but none of them seem to work well. At least you should keep it in mind that word embedding learns complicated linear relations between words.

I should explain word embedding techniques such as word2vec in detail, but the main focus of this article is not NLP, so the points I have mentioned are enough to understand Transformer model with NLP examples in the upcoming articles.


3 Language model

Language models is one of the most straightforward, but crucial ideas in NLP. This is also a big topic, so this article is going to cover only basic points. Language model is a mathematical model of the probabilities of which words to come next, given a context. For example if you have a sentence “In the lecture, he opened a _.”, a language model predicts what comes at the part “_.” It is obvious that this is contextual. If you are talking about general university students, “_” would be “textbook,” but if you are talking about Japanese universities, especially in liberal art department, “_” would be more likely to be “smartphone. I think most of you use this language model everyday. When you type in something on your computer or smartphone, you would constantly see text predictions, or they might even correct your spelling or grammatical errors. This is language modelling. You can make language models in several ways, such as n-gram and neural language models, but in this article I can explain only general formulations for such models.

*I am not sure which algorithm is used in which services. That must be too fast changing and competitive for me to catch up.

As I mentioned in the first article series on RNN, a sentence is usually processed as sequence data in NLP. One single sentence is denoted as \boldsymbol{X} = (\boldsymbol{x}^{(1)}, \dots, \boldsymbol{x}^{(\tau)}), a list of vectors. The vectors are usually embedding vectors, and the (t) is the index of the order of tokens. For example the sentence “You’ve go the power.” can be expressed as \boldsymbol{X} = (\boldsymbol{x}^{(1)}, \boldsymbol{x}^{(2)}, \boldsymbol{x}^{(3)}, \boldsymbol{x}^{(4)}), where \boldsymbol{x}^{(1)}, \boldsymbol{x}^{(2)}, \boldsymbol{x}^{(3)}, \boldsymbol{x}^{(4)} denote “You’ve”, “got”, “the”, “power”, “.” respectively. In this case \tau = 4.

In practice a sentence \boldsymbol{X} usually includes two tokens BOS and EOS at the beginning and the end of the sentence. They mean “Beginning Of Sentence” and “End Of Sentence” respectively. Thus in many cases \boldsymbol{X} = (\boldsymbol{BOS} , \boldsymbol{x}^{(1)}, \dots, \boldsymbol{x}^{(\tau)}, \boldsymbol{EOS} ). \boldsymbol{BOS} and \boldsymbol{EOS} are also both vectors, at least in the Tensorflow tutorial.

P(\boldsymbol{X} = (\boldsymbol{BOS}, \boldsymbol{x}^{(1)}, \dots, \boldsymbol{x}^{(\tau)}, \boldsymbol{EOS}) is the probability of incidence of the sentence. But it is easy to imagine that it would be very hard to directly calculate how likely the sentence \boldsymbol{X} appears out of all possible sentences. I would rather say it is impossible. Thus instead in NLP we calculate the probability P(\boldsymbol{X}) as a product of the probability of incidence or a certain word, given all the words so far. When you’ve got the words (\boldsymbol{x}^{(1)}, \dots, \boldsymbol{x}^{(t-1}) so far, the probability of the incidence of \boldsymbol{x}^{(t)}, given the context is  P(\boldsymbol{x}^{(t)}|\boldsymbol{x}^{(1)}, \dots, \boldsymbol{x}^{(t-1)}). P(\boldsymbol{BOS}) is a probability of the the sentence \boldsymbol{X} being (\boldsymbol{BOS}), and the probability of \boldsymbol{X} being (\boldsymbol{BOS}, \boldsymbol{x}^{(1)}) can be decomposed this way: P(\boldsymbol{BOS}, \boldsymbol{x}^{(1)}) = P(\boldsymbol{x}^{(1)}|\boldsymbol{BOS})P(\boldsymbol{BOS}).

Just as well P(\boldsymbol{BOS}, \boldsymbol{x}^{(1)}, \boldsymbol{x}^{(2)}) = P(\boldsymbol{x}^{(2)}| \boldsymbol{BOS}, \boldsymbol{x}^{(1)}) P( \boldsymbol{BOS}, \boldsymbol{x}^{(1)})= P(\boldsymbol{x}^{(2)}| \boldsymbol{BOS}, \boldsymbol{x}^{(1)}) P(\boldsymbol{x}^{(1)}| \boldsymbol{BOS}) P( \boldsymbol{BOS}).

Hence, the general probability of incidence of a sentence \boldsymbol{X} is P(\boldsymbol{X})=P(\boldsymbol{BOS}, \boldsymbol{x}^{(1)}, \boldsymbol{x}^{(2)}, \dots, \boldsymbol{x}^{(\tau -1)}, \boldsymbol{x}^{(\tau)}, \boldsymbol{EOS}) = P(\boldsymbol{EOS}| \boldsymbol{BOS}, \boldsymbol{x}^{(1)}, \dots, \boldsymbol{x}^{(\tau)}) P(\boldsymbol{x}^{(\tau)}| \boldsymbol{BOS}, \boldsymbol{x}^{(1)}, \dots, \boldsymbol{x}^{(\tau - 1)}) \cdots P(\boldsymbol{x}^{(2)}| \boldsymbol{BOS}, \boldsymbol{x}^{(1)}) P(\boldsymbol{x}^{(1)}| \boldsymbol{BOS}) P(\boldsymbol{BOS}).

Let \boldsymbol{x}^{(0)} be \boldsymbol{BOS} and \boldsymbol{x}^{(\tau + 1)} be \boldsymbol{EOS}. Plus, let P(\boldsymbol{x}^{(t+1)}|\boldsymbol{X}_{[0, t]}) be P(\boldsymbol{x}^{(t+1)}|\boldsymbol{x}^{(0)}, \dots, \boldsymbol{x}^{(t)}), then P(\boldsymbol{X}) = P(\boldsymbol{x}^{(0)})\prod_{t=0}^{\tau}{P(\boldsymbol{x}^{(t+1)}|\boldsymbol{X}_{[0, t]})}. Language models calculate which words to come sequentially in this way.

Here’s a question: how would you evaluate a language model?

I would say the answer is, when the language model generates words, the more confident the language model is, the better the language model is. Given a context, when the distribution of the next word is concentrated on a certain word, we can say the language model is confident about which word to come next, given the context.

*For some people, it would be more understandable to call this “entropy.”

Let’s take the vocabulary {“the”, “You’ve”, “Walberg”, “touch”, “power”, “Nights”, “got”, “Mark”, “Boogie”} as an example. Assume that P(\boldsymbol{X}) = P(\boldsymbol{BOS}, \boldsymbol{You've}, \boldsymbol{got}, \boldsymbol{the}, \boldsymbol{touch}, \boldsymbol{EOS}) = P(\boldsymbol{BOS}, \boldsymbol{x}^{(1)}, \boldsymbol{x}^{(2)}, \boldsymbol{x}^{(3)}, \boldsymbol{x}^{(4)}, \boldsymbol{EOS})= P(\boldsymbol{x}^{(0)})\prod_{t=0}^{4}{P(\boldsymbol{x}^{(t+1)}|\boldsymbol{X}_{[0, t]})}. Given a context (\boldsymbol{BOS}, \boldsymbol{x}^{(1)}), the probability of incidence of \boldsymbol{x}^{(2)} is P(\boldsymbol{x}^{2}|\boldsymbol{BOS}, \boldsymbol{x}^{(1)}). In the figure below, the distribution at the left side is less confident because probabilities do not spread widely, on the other hand the one at the right side is more confident that next word is “got” because the distribution concentrates on “got”.

*You have to keep it in mind that the sum of all possible probability P(\boldsymbol{x}^{(2)} | \boldsymbol{BOS}, \boldsymbol{x}^{(1)}) is 1, that is, P(\boldsymbol{the}| \boldsymbol{BOS}, \boldsymbol{x}^{(1)}) + P(\boldsymbol{You've}| \boldsymbol{BOS}, \boldsymbol{x}^{(1)}) + \cdots + P(\boldsymbol{Boogie}| \boldsymbol{BOS}, \boldsymbol{x}^{(1)}) = 1.

While the language model generating the sentence “BOS You’ve got the touch EOS”, it is better if the language model keeps being confident. If it is confident, P(\boldsymbol{X})= P(\boldsymbol{BOS}) P(\boldsymbol{x}^{(1)}|\boldsymbol{BOS}}P(\boldsymbol{x}^{(3)}|\boldsymbol{BOS}, \boldsymbol{x}^{(1)}, \boldsymbol{x}^{(2)}) P(\boldsymbol{x}^{(4)}|\boldsymbol{BOS}, \boldsymbol{x}^{(1)}, \boldsymbol{x}^{(2)}, \boldsymbol{x}^{(3)}) P(\boldsymbol{EOS}|\boldsymbol{BOS}, \boldsymbol{x}^{(1)}, \boldsymbol{x}^{(2)}, \boldsymbol{x}^{(3)}, \boldsymbol{x}^{(4)})} gets higher. Thus (-1) \{ log_{b}{P(\boldsymbol{BOS})} + log_{b}{P(\boldsymbol{x}^{(1)}|\boldsymbol{BOS}}) + log_{b}{P(\boldsymbol{x}^{(3)}|\boldsymbol{BOS}, \boldsymbol{x}^{(1)}, \boldsymbol{x}^{(2)})} + log_{b}{P(\boldsymbol{x}^{(4)}|\boldsymbol{BOS}, \boldsymbol{x}^{(1)}, \boldsymbol{x}^{(2)}, \boldsymbol{x}^{(3)})} + log_{b}{P(\boldsymbol{EOS}|\boldsymbol{BOS}, \boldsymbol{x}^{(1)}, \boldsymbol{x}^{(2)}, \boldsymbol{x}^{(3)}, \boldsymbol{x}^{(4)})} \} gets lower, where usually b=2 or b=e.

This is how to measure how confident language models are, and the indicator of the confidence is called perplexity. Assume that you have a data set for evaluation \mathcal{D} = (\boldsymbol{X}_1, \dots, \boldsymbol{X}_n, \dots, \boldsymbol{X}_{|\mathcal{D}|}), which is composed of |\mathcal{D}| sentences in total. Each sentence \boldsymbol{X}_n = (\boldsymbol{x}^{(0)})\prod_{t=0}^{\tau ^{(n)}}{P(\boldsymbol{x}_{n}^{(t+1)}|\boldsymbol{X}_{n, [0, t]})} has \tau^{(n)} tokens in total excluding \boldsymbol{BOS}, \boldsymbol{EOS}. And let |\mathcal{V}| be the size of the vocabulary of the language model. Then the perplexity of the language model is b^z, where z = \frac{-1}{|\mathcal{V}|}\sum_{n=1}^{|\mathcal{D}|}{\sum_{t=0}^{\tau ^{(n)}}{log_{b}P(\boldsymbol{x}_{n}^{(t+1)}|\boldsymbol{X}_{n, [0, t]})}. The b is usually 2 or e.

For example, assume that \mathcal{V} is vocabulary {“the”, “You’ve”, “Walberg”, “touch”, “power”, “Nights”, “got”, “Mark”, “Boogie”}. Also assume that the evaluation data set for perplexity of a language model is \mathcal{D} = (\boldsymbol{X}_1, \boldsymbol{X}_2), where \boldsymbol{X_1} =(\boldsymbol{You've}, \boldsymbol{got}, \boldsymbol{the}, \boldsymbol{touch}) \boldsymbol{X_2} = (\boldsymbol{You've}, \boldsymbol{got}, \boldsymbol{the }, \boldsymbol{power}). In this case |\mathcal{V}|=9, |\mathcal{D}|=2. I have already showed you how to calculate the perplexity of the sentence “You’ve got the touch.” above. You just need to do a similar thing on another sentence “You’ve got the power”, and then you can get the perplexity of the language model.

*If the network is not properly trained, it would also be confident of generating wrong outputs. However, such network still would give high perplexity because it is “confident” at any rate. I’m sorry I don’t know how to tackle the problem. Please let me put this aside, and let’s get down on Transformer model soon.


Let’s see how word embedding is implemented with a very simple example in the official Tensorflow tutorial. It is a simple binary classification task on IMDb Dataset. The dataset is composed to comments on movies by movie critics, and you have only to classify if the commentary is positive or negative about the movie. For example when you get you get an input “To be honest, Michael Bay is a terrible as an action film maker. You cannot understand what is going on during combat scenes, and his movies rely too much on advertisements. I got a headache when Mark Walberg used a Chinese cridit card in Texas. However he is very competent when it comes to humorous scenes. He is very talented as a comedy director, and I have to admit I laughed a lot.“, the neural netowork has to judge whether the statement is positive or negative.

This networks just takes an average of input embedding vectors and regress it into a one dimensional value from 0 to 1. The shape of embedding layer is (8185, 16). Weights of neural netowrks are usually implemented as matrices, and you can see that each row of the matrix corresponds to emmbedding vector of each token.

*It is easy to imagine that this technique is problematic. This network virtually taking a mean of input embedding vectors. That could mean if the input sentence includes relatively many tokens with negative meanings, it is inclined to be classified as negative. But for example, if the sentence is “This masterpiece is a dark comedy by Charlie Chaplin which depicted stupidity of the evil tyrant gaining power in the time. It thoroughly mocked Germany in the time as an absurd group of fanatics, but such propaganda could have never been made until ‘Casablanca.'” , this can be classified as negative, because only the part “masterpiece” is positive as a token, and there are much more words with negative meanings themselves.

The official Tensorflow tutorial provides visualization of word embedding with Embedding Projector, but I would like you to take more control over the data by yourself. Please just copy and paste the codes below, installing necessary libraries. You would get a map of vocabulary used in the text classification task. It seems you cannot find clear tendency of the clusters of the tokens. You can try other dimension reduction methods to get maps of the vocabulary by for example using Scikit Learn.


[1] “Word embeddings” Tensorflow Core

[2]Tsuboi Yuuta, Unno Yuuya, Suzuki Jun, “Machine Learning Professional Series: Natural Language Processing with Deep Learning,” (2017), pp. 43-64, 72-85, 91-94
坪井祐太、海野裕也、鈴木潤 著, 「機械学習プロフェッショナルシリーズ 深層学習による自然言語処理」, (2017), pp. 43-64, 72-85, 191-193

[3]”Stanford CS224N: NLP with Deep Learning | Winter 2019 | Lecture 8 – Translation, Seq2Seq, Attention”, stanfordonline, (2019)

[4] Francois Chollet, Deep Learning with Python,(2018), Manning , pp. 178-185

[5]”2.2. Manifold learning,” scikit-learn

* I make study materials on machine learning, sponsored by DATANOMIQ. I do my best to make my content as straightforward but as precise as possible. I include all of my reference sources. If you notice any mistakes in my materials, including grammatical errors, please let me know (email: And if you have any advice for making my materials more understandable to learners, I would appreciate hearing it.

Instructions on Transformer for people outside NLP field, but with examples of NLP

I found it quite difficult to explain mathematical details of long short-term memory (LSTM) in my previous article series. But when I was studying LSTM, a new promising algorithm was already attracting attentions. The algorithm is named Transformer. Its algorithm was a first announced in a paper named “Attention Is All You Need,” and it outperformed conventional translation algorithms with lower computational costs.

In this article series, I am going to provide explanations on minimum prerequisites for understanding deep learning in NLP (natural language process) tasks, but NLP is not the main focus of this article series, and actually I do not study in NLP field. I think Transformer is going to be a new major model of deep learning as well as CNN or RNN, and the model is now being applied in various fields.

Even though Transformer is going to be a very general deep learning model, I still believe it would be an effective way to understand Transformer with some NLP because language is a good topic we have in common. Unlike my previous article series, in which I tried to explain theoretical side of RNN as precisely as possible, in this article I am going to focus on practical stuff with my toy implementations of NLP tasks, largely based on Tensorflow official tutorial. But still I will do my best to make it as straightforward as possible to understand the architecture of Transformer with various original figures.

This series is going to be composed of the articles below.

If you are in the field and can read the codes in the official tutorial with no questions, this article series is not for you, but if you want to see how a Transformer works but do not want to go too much into details of NLP, this article would be for you.

Bias and Variance in Machine Learning

Machine learning continues to be an ever more vital component of our lives and ecosystem, whether we’re applying the techniques to answer research or business problems or in some cases even predicting the future. Machine learning models need to give accurate predictions in order to create real value for a given industry or domain.

While training a model is one of the key steps in the Data Science Project Life Cycle, how the model generalizes on unseen data is an equally important aspect that should be considered in every Data Science Project Life Cycle. We need to know whether it works and, consequently, if we can trust its predictions. Could the model be merely memorizing the data it is fed with, and therefore unable to make good predictions on future samples, or samples that it hasn’t seen before?

Let’s know the importance of evaluation with a simple example, There are two student’s Ramesh and Suresh preparing for the CAT exam to get into top IIMs (Indian Institute of Management). They both are quite good friends and stayed in the room during preparation and put an equal amount of hard work while solving numerical problems.

They both prepared for almost the same number of hours for the entire year and appeared in the final CAT exam. Surprisingly, Ramesh cleared, but Suresh did not. When asked, we got to know that there was one difference in their strategy of preparation between them, Ramesh had joined a Test Series course where he used to test his knowledge and understanding by giving mock exams and then further evaluating on which portions he is lagging and making necessary adjustments to he is preparation cycle in order to do well in those areas. But Suresh was confident, and he just kept training himself without testing on the preparation he had done.

Like the above situation we can train a Machine Learning Algorithm extensively with many parameters and new techniques, but if you are skipping its evaluation step, you cannot trust your model to perform well on the unseen data. In this article, we explain the importance of Bias, Variance and the trade-off between them in order to know how well a machine learning model generalizes to new, previously unseen data.

Training of Supervised Machine Learning


Bias is the difference between the Predicted Value and the Expected Value or how far are the predicted values from the actual values. During the training process the model makes certain assumptions on the training data provided. After Training, when it is introduced to the testing/validation data or unseen data, these assumptions may not always be correct.

If we use a large number of nearest neighbors in the K-Nearest Neighbors Algorithm, the model can totally decide that some parameters are not important at all for the modelling.  For example, it can just consider that only two predictor variables are enough to classify the data point though we have more than 10 variables.

This type of model will make very strong assumptions about the other parameters not affecting the outcome at all. You can take it as a model predicting or understanding only the simple relationship when the data points clearly indicate a more complex relationship.

When the model has high bias error, it results in a very simplistic model that does not consider the complexity of the data very well leading to Underfitting.


Variance occurs when the model performs well on the trained dataset but does not do well on an unseen data set, it is when the model considers the fluctuations or i.e. the noise as in the data as well. The model will still consider the variance as something to learn from because it learns too much from the noise inside the trained data set that it fails to perform as expected on the unseen data.

Based on the above example from Bias, if the model learns that all the ten predictor variables are important to classify a given data point then it tends to have high variance. You can take it as the model is trying to understand every minute detail making it more complex and failing to perform well on the unseen data.

When a model has High Bias error, it underfits the data and makes very simplistic assumptions on it. When a model has High Variance error, it overfits the data and learns too much from it. When a model has balanced Bias and Variance errors, it performs well on the unseen data.

Bias-Variance Trade-off

Based on the definitions of bias and variance, there is clear trade-off between bias and variance when it comes to the performance of the model. A model will have a high error if it has very high bias and low variance and have a high error if it has high variance and low bias.

A model that strikes a balance between the bias and variance can minimize the error better than those that live on extreme ends.

We can find whether the model has High Bias using the below steps:

  1. We tend to get high training errors.
  2. The validation error or test error will be similar to the training error.

We can find whether the model has High Bias using the below steps:

  1. We tend to get low training error
  2. The validation error or test error will be very high.

We can fix the High Bias using below steps:

  1. We need to gather more input features or can even try to create few using the feature engineering techniques.
  2. We can even add few polynomial features in order to increase the complexity.
  3. If we are using any regularization terms in our model, we can try to minimize it.

We can fix the High Variance using below steps:

  1. We can gather more training data so that the model can learn more on the patterns rather than the noise.
  2. We can even try to reduce the input features or do feature selection.
  3.  If we are using any regularization terms in our model we can try to maximize it.


In this article, we got to know the importance of the evaluation step in the Data Science Project Life Cycle, definitions of Bias and Variance, the trade-off between them and the steps we can take to fix the Underfitting and Overfitting of a Machine Learning Model.

Spiky cubes, Pac-Man walking, empty M&M’s chocolate: curse of dimensionality

This is the first article of the article series Illustrative introductions on dimension reduction.

“Curse of dimensionality” means the difficulties of machine learning which arise when the dimension of data is higher. In short if the data have too many features like “weight,” “height,” “width,” “strength,” “temperature”…., that can undermine the performances of machine learning. The fact might be contrary to your image which you get from the terms “big” data or “deep” learning. You might assume that the more hints you have, the better the performances of machine learning are. There are some reasons for curse of dimensionality, and in this article I am going to introduce two major reasons below.

  1. High dimensional data usually have rich expressiveness, but usually training data are too poor for that.
  2. The behaviors of data points in high dimensional space are totally different from our common sense.

Through these topics, you will see that you always have to think about which features to use considering the number of data points.

*From now on I am going to talk about only Euclidean distance. If you are not sure what Euclidean distance means, please just keep it in mind that it is the type of distance most people wold have learnt in normal compulsory education.

*This is the first article of the article series ” Illustrative introductions on dimension reduction .”

1. Number of samples and degree of dimension

The most straightforward demerit of adding many features, or increasing dimensions of data, is the growth of computational costs. More importantly, however, you always have to think about the degree of dimensions in relation of the number of data points you have. Let me take a simple example in a book “Pattern Recognition and Machine Learning” by C. M. Bishop (PRML). This is an example of measurements of a pipeline. The figure below shows a comparison plot of 3 classes (red, green and blue), with parameter x_7 plotted against parameter x_6 out of 12 parameters.

* The meaning of data is not important in this article. If you are interested please refer to the appendix in PRML.

Assume that we are interested in classifying the cross in black into one of the three classes. One of the most naive ideas of this classification is dividing the graph into grids and labeling each grid depending on the number of samples in the classes (which are colored at the right side of the figure). And you can classify the test sample, the cross in black, into the class of the grid where the test sample is in. Thereby the cross is classified to the class in red.

Source: C.M. Bishop, “Pattern Recognition and Machine Learning,” (2006), Springer, pp. 34-35

As I mentioned in the figure above, we used only two features out of 12 features in total. When the total number of data points is fixed and you add remaining ten axes/features one after another, what would happen? Let’s see what “adding axes/features” means. If you are talking about 1, 2, or 3 dimensional grids, you can visualize them. And as you can see from the figure below, if you make each 10^1, 10^2, 100^3 grids respectively in 1, 2, 3 dimensional spaces, the number of the small regions in the grids are respectively 10, 100, 1000. Even though you cannot visualize it anymore, you can make grids for more than 3 dimensional data. If you continue increasing the degree of dimension, the number of grids increases exponentially, and that can soon surpass the number of training data points. That means there would be a lot of empty spaces in such high dimensional grids. And the classifying method above: coloring each grid and classifying unknown samples depending on the colors of the grids, does not work out anymore because there would be a lot of empty grids.

* If you are still puzzled by the idea of “more than 3 dimensional grids,” you should not think too much about that now. It is enough if you can get some understandings on high dimensional data after reading the whole article of this.

Source: Goodfellow and Yoshua Bengio and Aaron Courville, Deep Learning, (2016), MIT Press, p. 153

I said the method above is the most naive way, but other classical classification methods , for example k-nearest neighbors algorithm, are more or less base on a similar idea. Many of classical machine learning algorithms are based on the idea of smoothness prior, or local constancy prior. In short in classical ways, you  do not expect data to change so much in a small region, so you can expect unknown samples to be similar to data in vicinity. But that soon turns out to be problematic when the dimension of data is bigger because training data would be sparse because the area of multidimensional space grows exponentially as I mentioned above. And sometimes you would not be able to find training data around test data. Plus, in high dimensional data, you cannot treat distance in the same as you do in lower dimensional space. The ideas of “close,” “nearby,” or “vicinity” get more obscure in high dimensional data. That point is related to the next topic: the intuition have cultivated in normal life is not applicable to higher dimensional data.

2. Bizarre characteristics of high dimensional data

We form our sense of recognition in 3-dimensional ways in our normal life. Even though we can visualize only 1, 2, or 3 dimensional data, we can actually generalize the ideas in 1, 2, or 3 dimensional ideas to higher dimensions. For example 4 dimensional cubes, 100 dimensional spheres, or orthogonality in 255 dimensional space. Again, you cannot exactly visualize those ideas, and for many people, such high dimensional phenomenon are just imaginary matters on blackboards. Those high dimensional ideas are designed to retain some conditions just as well as 1, 2, or 3 dimensional space. Let’s take an example of spheres in several dimensional spaces. General spheres in any D-dimensional space can be defined as a set of any \boldsymbol{x}, such that |\boldsymbol{x} - \boldsymbol{c}| = r, where \boldsymbol{c} is the center point and r is length of radius. When \boldsymbol{x} is 2-dimensional, the spheres are called “circles.” When \boldsymbol{x} is 3-dimensional, the spheres are called “spheres” in our normal life, unless it is used in a conversation in a college cafeteria, by some students in mathematics department. And when \boldsymbol{x} is D-dimensional, they are called D-ball, and again, this is just a imaginary phenomenon on blackboard.

* Vectors and points are almost the same because all the vectors are denoted as “arrows” from the an origin point to sample data points.  The only difference is that when you use vectors, you have to consider their directions.

* “D-ball” is usually called “n-ball,” and in such context it is a sphere in a n-dimensional space. But please let me use the term “D-ball” in this article.

Not only spheres, but only many other ideas have been generalized to D-dimensional space, and many of them are indispensable also for data science. But there is one severe problem: the behaviors of data in high dimensional field is quite different from those in two or three dimensional space. To be concrete, in high dimensional field, cubes are spiky, you have to move like Pac-Man, and M & M’s Chocolate looks empty inside but tastes normal.

2.1: spiky cubes
Let’s take a look at an elementary-school-level example of geometry first. Assume that you have several unit squares or unit cubes like below. In each of them a circle or sphere with diameter 1 is inscribed. The length of a diagonal line in each square is \sqrt{2}, and that in each cube is \sqrt{3}.

If you stack the squares or cubes as below, what are the length of diameters of the blue circle or sphere, circumscribing all the 4 orange circles or the 8 orange spheres?

The answers are, the diameter of the blue circle is \sqrt{2} - 1, and the diameter of the blue sphere is \sqrt{3} - 1.

Next let’s think about the same situation in higher dimensional space. Assume that there are some unit D-dimensional hypercubes stacked, in each of which a D-ball with diameter 1 is inscribed, touching all the surfaces inside. Then what is the length of the diameter of  a D-ball circumscribing all the unit D-ball in the hypercubes ? Given the results above, it ca be predicted that its diameter is \sqrt{D}  -1. If that is true, there is one strange point: \sqrt{D} - 1 can soon surpass 2: that means in the chart above the blue sphere will stick out of the stacked cubes. That sounds like a paradox, but with one hypothesis, the phenomenon makes sense: cubes become more spiky as the degree of dimension grows. This hypothesis is a natural deduction because diagonal lines of hyper cubes get longer, and the the center of each surface of hypercubes still touches the unit D-ball with diameter 1, inscribing inscribing inside each unit hypercube.

If you stack 4 hypercubes, the blue sphere circumscribing them will not stick out of the stacked hypercubes anymore like the figure below.

*Of course you cannot visualize what is going on in D-dimensional space, so the figure below is just a pseudo simulation of D-dimensional space in our 3-dimensional sense. I guess you have to stack more than four hyper cubes in higher dimensional data, but you cannot easily imagine what will go on in such space anymore.


*You can confirm the fact that hypercube gets more spiky as the degree of dimension growth, by comparing the volume of the hypercube and the volume of the D-ball inscribed inside the hypercube. Thereby you can prove that the volume of hypercube concentrates on the corners of the hypercube. Plus, as I mentioned the longest diagonal distance of hypercube gets longer as dimension degree increases. That is why hypercube is said to be spiky. For mathematical proof, please check the Exercise 1.19 of PRML.

2.2: Pac-Man walking

Next intriguing phenomenon in high dimensional field is that most of pairs of vectors in high dimensional space are orthogonal. In other words, if you select two random vectors in high dimensional space, the angle between them are mostly close to 90^\circ. Let’s see the general meaning of angle between two vectors in any dimensional spaces. Assume that the angle between two vectors \boldsymbol{u}, and \boldsymbol{v} is \theta, then cos\theta is calculated as cos\theta = \frac{<\boldsymbol{u}, \boldsymbol{v}>}{|\boldsymbol{u}||\boldsymbol{v}|}. In 1, 2, or 3 dimensional space, you can actually see the angle, but again you can define higher dimensional angle, which you cannot visualize anymore. And angles are sometimes used as similarity of two vectors.

* <\boldsymbol{u}, \boldsymbol{v}> is the inner product of \boldsymbol{u}, and \boldsymbol{v}.

Assume that you generate a pair of two points inside a D-dimensional unit sphere and make two vectors \boldsymbol{u}, and \boldsymbol{v} by connecting the origin point and those two points respectively. When D is 2, I mean spheres are circles in this case, any \theta are equally generated as in the chart below. The fact might be the same as your intuition.   How about in 3-dimensional space? In fact the distribution of \theta is not uniform. \theta = 90^\circ is the most likely to be generated. As I explain in the figure below, if you compare the area of cross section of a hemisphere and the area of a cone whose vertex is the center point of the sphere, you can see why.

I generated 10000 random pairs of points in side a D-dimensional unit sphere, and calculated the angle between them. In other words I just randomly generated two D-dimensional vectors \boldsymbol{u} and \boldsymbol{v}, whose elements are randomly generated values between -1 and 1, and calculated the angle between them, repeating this process 10000 times. The chart below are the histograms of angle between pairs of generated vectors in respectively 2, 3, 50, and 100 dimensional space.

As I explained above, in 2-dimensional space, the distribution of \theta is almost uniform. However the distribution concentrates a little around 90^\circ in 3-dimensional space. You can see that the bigger the degree of dimension is, the more the angles of generated vectors concentrate around 90^\circ. That means most pairs of vectors in high dimensional space are close to orthogonal. Movements are also sequence of vectors, so when most pairs of movement vectors are orthogonal, that means you can only move like Pac-Man in such space.


* Of course I am talking about arcade Mac-Man game. Not Pac-Man in Super Smash Bros.  Retro RPG video games might have more similar playability, but in high dimensional space it is also difficult to turn back. At any rate, I think you have understood it is even difficult to move smoothly in high dimensional space, just like the first notorious Resident Evil on the first PS console also had terrible playability .

2.3: empty M & M’s chocolate

Let’s think about the proportion of the volume of the outermost \epsilon surface of general spheres with radius r. First, in 2 two dimensional space, spheres are circles. The area of the brown part of the circle below is \pi r^2. In order calculate the are of \epsilon \cdot r thick surface of the circle, you have only to subtract the area of \pi \{ (1 - \epsilon)\cdot r\} ^2. When \epsilon = 0.01, the area of outer most surface is \pi r^2 - \pi (0.99\cdot r)^2, and its proportion to the area of the whole circle is \frac{\pi r^2 - \pi (0.99\cdot r)^2}{\pi r^2} = 0.0199.

In case of 3-dimensional space, the value of a sphere with radius r is \frac{4}{3} \pi r^2, so the proportion of the \epsilon surface is calculated in the same way: \frac{\frac{4}{3} \pi r^3 -\frac{4}{3} \pi (0.99\cdot r)^2}{\frac{4}{3}\pi r^2} = 0.0297. Compared to the case in 2 dimensional space, the proportion is a little bigger.

How about in D-dimensional space? We have seen that even in  D-dimensional space the surface of a sphere, I mean D-ball, can be defined as a set of any points whose distance from the center point is all r. And it is known that the volume of D-ball is defined as below.

\Gamma () is called gamma function, but in this article it is not so important. The most important point now is, if you discuss any D-ball, their volume only depends on their radius r. That meas the proportion of outer \epsilon surface of D-ball is calculated as \frac{\pi r^2 - \pi \{ (1 - \epsilon)\cdot r\} ^2}{\pi r^2}. When \epsilon is 0.01, the proportion of the 1% surface of D-ball changes like in the chart below.

* And of course when D is 2,  \frac{\pi ^{(\frac{D}{2})}}{\Gamma (\frac{D}{2} + 1)} = \pi, and when D is 3 ,  \frac{\pi ^{(\frac{D}{2})}}{\Gamma (\frac{D}{2} + 1)} = \frac{4}{3} \pi

You can see that when D is over 400, around 90% of volume is concentrated in the very thin 1% surface of D-ball. That is why, in high dimensional space, M & M’s chocolate look empty but tastes normal: all the chocolate are concentrated beneath the sugar coating.

More interestingly, even if you choose any points as a central point of a sphere with radius r, the other points are squashed to the surface of the sphere, even if all the data points are uniformly distributed. This situation is problematic for classical machine learning algorithms, which are often based on the Euclidean distances between pairs of two sample data points: if you go from the central point to another sample point, the possibility of finding the point within (1 - \epsilon)\cdot r radius of the center is almost zero. But if you reach the outermost \epsilon part of the surface of the sphere, most data points are there. However, for one of the data points in the surface, any other data points are distant in the same way.

Inside M & M’s chocolate is a mysterious world.


You have seen that using high dimensional data can be problematic in many ways. Data science and machine learning are largely based on one idea: you can find a lower dimensional meaningful and easier structure in data. In the next articles I am going to introduce some famous dimension reduction algorithms. And hopefully I would like to give some deeper insights in to these algorithms, in straightforward ways.

* I could not explain the relationships of variance and bias of data. This is also a very important factor when you think about dimensionality of data. I hope I can write about this topic someday. You can also look it up if you are interested.


[1]C. M. Bishop, “Pattern Recognition and Machine Learning,” (2006), Springer, pp. 33-37

[2]Goodfellow and Yoshua Bengio and Aaron Courville, Deep Learning, (2016), MIT Press, p. 153

[3] Shiga Kouji, “30 Lesson to Topology,” (1988)

[4]”Volume of an n-ball,” Wikipedia

* I make study materials on machine learning, sponsored by DATANOMIQ. I do my best to make my content as straightforward but as precise as possible. I include all of my reference sources. If you notice any mistakes in my materials, including grammatical errors, please let me know (email: And if you have any advice for making my materials more understandable to learners, I would appreciate hearing it.

Simple RNN

LSTM back propagation: following the flows of variables

First of all, the summary of this article is: please just download my Power Point slides which I made and be patient, following the equations.

I am not supposed to use so many mathematics when I write articles on Data Science Blog. However using little mathematics when I talk about LSTM backprop is like writing German, never caring about “der,” “die,” “das,” or speaking little English in English classes (which most high school English teachers in Japan do) or writing Japanese without using any Chinese characters (which looks like a terrible handwriting by a drug addict). In short, that is ridiculous. And all the precise equations of LSTM backprop, written on a blog is not a comfortable thing to see. So basically the whole of this article is an advertisement on my PowerPoint slides, sponsored by DATANOMIQ, and I can just give you some tips to get ready for the most tiresome part of understanding LSTM here.

*This article is the fifth article of “A gentle introduction to the tiresome part of understanding RNN.”

 *In this article “Densely Connected Layers” is written as “DCL,” and “Convolutional Neural Network” as “CNN.”

1. Chain rules

This article is virtually an article on chain rules of differentiation. Even if you have clear understandings on chain rules, I recommend you to take a look at this section. If you have written down all the equations of back propagation of DCL, you would have seen what chain rules are. Even simple chain rules for backprop of normal DCL can be difficult to some people, but when it comes to backprop of LSTM, it is a pure torture.  I think using graphical models would help you understand what chain rules are like. Graphical models are basically used to describe the relations of variables and functions in probabilistic models, so to be exact I am going to use “something like graphical models” in this article. Not that this is a common way to explain chain rules.

First, let’s think about the simplest type of chain rule. Assume that you have a function f=f(x)=f(x(y)), and relations of the functions are displayed as the graphical model at the left side of the figure below. Variables are a type of function, so you should think that every node in graphical models denotes a function. Arrows in purple in the right side of the chart show how information propagate in differentiation.

Next, if you have a function f , which has two variances  x_1 and x_2. And both of the variances also share two variances  y_1 and y_2. When you take partial differentiation of f with respect to y_1 or y_2, the formula is a little tricky. Let’s think about how to calculate \frac{\partial f}{\partial y_1}. The variance y_1 propagates to f via x_1 and x_2. In this case the partial differentiation has two terms as below.

In chain rules, you have to think about all the routes where a variance can propagate through. If you generalize chain rules as the graphical model below, the partial differentiation of f with respect to y_i is calculated as below. And you need to understand chain rules in this way to understanding any types of back propagation.

The figure above shows that if you calculate partial differentiation of f with respect to y_i, the partial differentiation has n terms in total because y_i propagates to f via n variances. In order to understand backprop of LSTM, you constantly have to care about the flows of variances, which I display as purple arrows.

2. Chain rules in LSTM

I would like you to remember the figure below, which I used in the second article to show how errors propagate backward during backprop of simple RNNs. After forward propagation, first of all, you need to calculate \frac{\partial J}{\partial \boldsymbol{\theta}^{(t)}}, gradients of the error function with respect to parameters, at each time step. But you have to be careful that even though these gradients depend on time steps, the parameters \boldsymbol{\theta} do not depend on time steps.

*As I mentioned in the second article I personally think \frac{\partial J}{\partial \boldsymbol{\theta}^{(t)}} should be rather denoted as (\frac{\partial J}{\partial \boldsymbol{\theta}})^{(t)} because parameters themselves do not depend on time. However even the textbook by MIT press partly use the former notation. And I think you are likely to encounter this type of notation, so I think it is not bad to get ready for both.

The errors at time step (t) propagate backward to all the \boldsymbol{h} ^{(s)} (s \leq t). Conversely, in order to calculate \frac{\partial J}{\partial \boldsymbol{\theta}^{(t)}} errors flowing from J^{(s)}  (s \geq t). In the chart you need arrows of errors in purple for the gradient in a purple frame, orange arrows for gradients in orange frame, red arrows for gradients in red frame. And you need to sum up \frac{\partial J}{\partial \boldsymbol{\theta}^{(t)}} to calculate \frac{\partial J}{\partial \boldsymbol{\theta}} = \sum_{t}{\frac{\partial J}{\partial \boldsymbol{\theta}^{(t)}}}, and you need this gradient \frac{\partial J}{\partial \boldsymbol{\theta}} to renew parameters, one time.

At an RNN block level, the flows of errors and how to renew parameters are the same in LSTM backprop, but the flow of errors inside each block is much more complicated in LSTM backprop. But in order to denote errors of LSTM backprop, instead of \frac{\partial J}{\partial \boldsymbol{\theta}^{(t)}}, I use a special notation \delta \star ^{(t)} = \frac{\partial J}{\partial \star}.

* Again, please be careful of what \delta \star  ^{(t)} means. Neurons depend on time steps, but parameters do not depend on time steps. So if \star are neurons,  \delta \star  ^{(t)}= \frac{\partial J}{ \partial \star ^{(t)}}, but when \star are parameters, \delta \star  ^{(t)} should be rather denoted like \delta \star  ^{(t)}= (\frac{\partial J}{ \partial \star })^{(t)}. In the Space Odyssey paper\boldsymbol{\star} are not used as parameters, but in my PowerPoint slides and some other materials, \boldsymbol{\star} are used also as parameteres.

As I wrote in the last article, you calculate \boldsymbol{f}^{(t)}, \boldsymbol{i}^{(t)}, \boldsymbol{z}^{(t)}, \boldsymbol{o}^{(t)} as below. Unlike the last article, I also added the terms of peephole connections in the equations below, and I also introduced the variances \bar{\boldsymbol{f}}^{(t)}, \bar{\boldsymbol{i}}^{(t)}, \bar{\boldsymbol{z}}^{(t)}, \bar{\boldsymbol{o}}^{(t)} for convenience.

  • \boldsymbol{\bar{f}}^{(t)}=\boldsymbol{W}_{for} \cdot \boldsymbol{x}^{(t)} + \boldsymbol{R}_{for} \cdot \boldsymbol{y}^{(t-1)} + \boldsymbol{p}_{for}\odot \boldsymbol{c}^{(t-1)} + \boldsymbol{b}_{for}
  • \boldsymbol{\bar{i}}^{(t)}=\boldsymbol{W}_{in} \cdot \boldsymbol{x}^{(t)} + \boldsymbol{R}_{in} \cdot \boldsymbol{y}^{(t-1)} + \boldsymbol{p}_{in}\odot \boldsymbol{c}^{(t-1)} + \boldsymbol{b}_{in}
  • \boldsymbol{\bar{z}}^{(t)}=\boldsymbol{W}_z \cdot \boldsymbol{x}^{(t)} + \boldsymbol{R}_z \cdot \boldsymbol{y}^{(t-1)} + \boldsymbol{b}_z
  • \boldsymbol{\bar{o}}^{(t)}=\boldsymbol{W}_{out} \cdot \boldsymbol{x}^{(t)} + \boldsymbol{R}_{out} \cdot \boldsymbol{y}^{(t-1)} + \boldsymbol{p}_{out}\odot \boldsymbol{c}^{(t)} + \boldsymbol{b}_{out}
  • \boldsymbol{f}^{(t)}=\sigma( \boldsymbol{\bar{f}}^{(t)})
  • \boldsymbol{i}^{(t)}=\sigma(\boldsymbol{\bar{i}}^{(t)})
  • \boldsymbol{z}^{(t)}=tanh(\boldsymbol{\bar{z}}^{(t)})
  • \boldsymbol{o}^{(t)}=\sigma(\boldsymbol{\bar{o}}^{(t)})

With  Hadamar product operator, the renewed cell and the output are calculated as below.

  • \boldsymbol{c}^{(t)} = \boldsymbol{z}^{(t)}\odot \boldsymbol{i}^{(t)} + \boldsymbol{c}^{(t-1)} \odot \boldsymbol{f}^{(t)}
  • \boldsymbol{y}^{(t)} = \boldsymbol{o}^{(t)} \odot tanh(\boldsymbol{c}^{(t)})

In this article I would rather give instructions on how to read my PowerPoint slide. Just as general backprop, you need to calculate gradients of error functions with respect to parameters, such as \delta \boldsymbol{W}_{\star}, \delta \boldsymbol{R}_{\star}, \delta \boldsymbol{p}_{\star}, \delta \boldsymbol{b}_{\star}, where \star is either of \{z, in, for, out \}. And just as backprop of simple RNNs, in order to calculate gradients with respect to parameters, you need to calculate errors of neurons, that is gradients of error functions with respect to neurons, such as \delta \boldsymbol{f}^{(t)}, \delta \boldsymbol{i}^{(t)}, \delta \boldsymbol{z}^{(t)}, \delta \boldsymbol{o}^{(t)}.

*Again and again, keep it in mind that neurons depend on time steps, but parameters do not depend on time steps.

When you calculate gradients with respect to neurons, you can first calculate \delta \boldsymbol{y}^{(t)}, but the equation for this error is the most difficult, so I recommend you to put it aside for now. After calculating \delta \boldsymbol{y}^{(t)}, you can next calculate \delta \bar{\boldsymbol{o}}^{(t)}= \frac{\partial J^{(t)}}{ \partial \bar{\boldsymbol{o}}^{(t)}}. If you see the LSTM block below as a graphical model which I introduced, the information of \bar{\boldsymbol{o}}^{(t)} flow like the purple arrows. That means, \bar{\boldsymbol{o}}^{(t)} affects J only via \boldsymbol{y}^{(t)}, and this structure is equal to the first graphical model which I have introduced above. And if you calculate \bar{\boldsymbol{o}}^{(t)} element-wise, you get the equation \delta \bar{o}_{k}^{(t)}=\frac{\partial J}{\partial \bar{o}_{k}^{(t)}}= \frac{\partial J}{\partial y_{k}^{(t)}} \frac{\partial y_{k}^{(t)}}{\partial \bar{o}_{k}^{(t)}}.

*The k is an index of an element of vectors. If you can calculate element-wise gradients, it is easy to understand that as differentiation of vectors and matrices.

Next you can calculate \delta \boldsymbol{c}^{(t)}, and chain rules are very important in this process. The flow of \delta \boldsymbol{c}^{(t)} to J can be roughly divided into two streams: the one which flows to J as \bodlsymbol{y}^{(t)}, and the one which flows to J as \bodlsymbol{c}^{(t+1)}. And the stream from \bodlsymbol{c}^{(t)} to \bodlsymbol{y}^{(t)} also have two branches: the one via \bar{\boldsymbol{o}}^{(t)} and the one which directly converges as  \bodlsymbol{y}^{(t)}. Just as well, the stream from \bodlsymbol{c}^{(t)} to \bodlsymbol{c}^{(t+1)} also have three branches: the ones via \bar{\boldsymbol{f}}^{(t)}, \bar{\boldsymbol{i}}^{(t)}, and the one which directly converges as \bodlsymbol{c}^{(t+1)}.

If you see see these flows as graphical a graphical model, that would be like the figure below.

According to this graphical model, you can calculate \delta \boldsymbol{c} ^{(t)} element-wise as below.

* TO BE VERY HONEST I still do not fully understand why we can apply chain rules like above to calculate \delta \boldsymbol{c}^{(t)}. When you apply the formula of chain rules, which I showed in the first section, to this case, you have to be careful of where to apply partial differential operators \frac{\partial}{ \partial c_{k}^{(t)}}. In the case above, in the part \frac{\partial y_{k}^{(t)}}{\partial c_{k}^{(t)}} the partial differential operator only affects tanh(c_{k}^{(t)}) of o_{k}^{(t)} \cdot tanh(c_{k}^{(t)}). And in the part \frac{\partial c_{k}^{(t+1)}}{\partial c_{k}^{(t)}}, the partial differential operator \frac{\partial}{\partial c_{k}^{(t)}} only affects the part c_{k}^{(t)} of the term c^{t}_{k} \cdot f_{k}^{(t+1)}. In the \frac{\partial \bar{o}_{k}^{(t)}}{\partial c_{k}^{(t)}} part, only (p_{out})_{k} \cdot c_{k}^{(t)},  in the \frac{\partial \bar{i}_{k}^{(t+1)}}{\partial c_{k}^{(t)}} part, only (p_{in})_{k} \cdot c_{k}^{(t)}, and in the \frac{\partial \bar{f}_{k}^{(t+1)}}{\partial c_{k}^{(t)}} part, only (p_{in})_{k} \cdot c_{k}^{(t)}. But some other parts, which are not affected by \frac{\partial}{ \partial c_{k}^{(t)}} are also functions of c_{k}^{(t)}. For example o_{k}^{(t)} of o_{k}^{(t)} \cdot tanh(c_{k}^{(t)}) is also a function of c_{k}^{(t)}. And I am still not sure about the logic behind where to affect those partial differential operators.

*But at least, these are the only decent equations for LSTM backprop which I could find, and a frequently cited paper on LSTM uses implementation based on these equations. Computer science is more of practical skills, rather than rigid mathematical logic. Also I think I have spent great deal of my time thinking about this part, and it is now time for me to move to next step. If you have any comments or advice on this point, please let me know.

Calculating \delta \bar{\boldsymbol{f}}^{(t)}, \delta \bar{\boldsymbol{i}}^{(t)}, \delta \bar{\boldsymbol{z}}^{(t)} are also relatively straigtforward as calculating \delta \bar{\boldsymbol{o}}^{(t)}. They all use the first type of chain rule in the first section. Thereby you can get these gradients: \delta \bar{f}_{k}^{(t)}=\frac{\partial J}{ \partial \bar{f}_{k}^{(t)}} =\frac{\partial J}{\partial c_{k}^{(t)}} \frac{\partial c_{k}^{(t)}}{ \partial \bar{f}_{k}^{(t)}}, \delta \bar{i}_{k}^{(t)}=\frac{\partial J}{\partial \bar{i}_{k}^{(t)}} =\frac{\partial J}{\partial c_{k}^{(t)}} \frac{\partial c_{k}^{(t)}}{ \partial \bar{i}_{k}^{(t)}}, and \delta \bar{z}_{k}^{(t)}=\frac{\partial J}{\partial \bar{z}_{k}^{(t)}} =\frac{\partial J}{\partial c_{k}^{(t)}} \frac{\partial c_{k}^{(t)}}{ \partial \bar{i}_{k}^{(t)}}.

All the gradients which we have calculated use the error \delta \boldsymbol{y}^{(t)}, but when it comes to calculating \delta \boldsymbol{y}^{(t)}….. I can only say “Please be patient. I did my best in my PowerPoint slides to explain that.” It is not a kind of process which I want to explain on Word Press. In conclusion you get an error like this: \delta \boldsymbol{y}^{(t)}=\frac{\partial J^{(t)}}{\partial \boldsymbol{y}^{(t)}} + \boldsymbol{R}_{for}^{T} \delta \bar{\boldsymbol{f}}^{(t+1)} + \boldsymbol{R}_{in}^{T}\delta \bar{\boldsymbol{i}}^{(t+1)} + \boldsymbol{R}_{out}^{T}\delta \bar{\boldsymbol{o}}^{(t+1)} + \boldsymbol{R}_{z}^{T}\delta \bar{\boldsymbol{z}}^{(t+1)}, and the flows of \boldsymbol{y}^{(t)} are as blow.

Combining the gradients we have got so far, we can calculate gradients with respect to parameters. For concrete results, please check the Space Odyssey paper or my PowerPoint slide.

3. How LSTMs tackle exploding/vanishing gradients problems

*If you are allergic to mathematics, you should not read this section or even download my PowerPoint slide.

*Part of this section is more or less subjective, so if you really want to know how LSTM mitigate the problems, I highly recommend you to also refer to other materials. But at least I did my best for this article.

LSTMs do not completely solve, vanishing gradient problems. They mitigate vanishing/exploding gradient problems. I am going to roughly explain why they can tackle those problems. I think you find many explanations on that topic, but many of them seems to have some mathematical mistakes (even the slide used in a lecture in Stanford University) and I could not partly agree with some statements. I also could not find any papers or materials which show the whole picture of how LSTMs can tackle those problems. So in this article I am only going to give instructions on the major way to explain this topic.

First let’s see how gradients actually “vanish” or “explode” in simple RNNs. As I in the second article of this series, simple RNNs propagate forward as the equations below.

  • \boldsymbol{a}^{(t)} = \boldsymbol{b} + \boldsymbol{W} \cdot \boldsymbol{h}^{(t-1)} + \boldsymbol{U} \cdot \boldsymbol{x}^{(t)}
  • \boldsymbol{h}^{(t)}= g(\boldsymbol{a}^{(t)})
  • \boldsymbol{o}^{(t)} = \boldsymbol{c} + \boldsymbol{V} \cdot \boldsymbol{h}^{(t)}
  • \hat{\boldsymbol{y}} ^{(t)} = f(\boldsymbol{o}^{(t)})

And every time step, you get an error function J^{(t)}. Let’s consider the gradient of J^{(t)} with respect to \boldsymbol{h}^{(k)}, that is the error flowing from J^{(t)} to \boldsymbol{h}^{(k)}. This error is the most used to calculate gradients of the parameters in the equations above.

If you calculate this error more concretely, \frac{\partial J^{(t)}}{\partial \boldsymbol{h}^{(k)}} = \frac{\partial J^{(t)}}{\partial \boldsymbol{h}^{(t)}} \frac{\partial \boldsymbol{h}^{(t)}}{\partial \boldsymbol{h}^{(t-1)}} \cdots \frac{\partial \boldsymbol{h}^{(k+2)}}{\partial \boldsymbol{h}^{(k+1)}} \frac{\partial \boldsymbol{h}^{(k+1)}}{\partial \boldsymbol{h}^{(k)}} = \frac{\partial J^{(t)}}{\partial \boldsymbol{h}^{(t)}} \prod_{k< s \leq t} \frac{\partial \boldsymbol{h}^{(s)}}{\partial \boldsymbol{h}^{(s-1)}}, where \frac{\partial \boldsymbol{h}^{(s)}}{\partial \boldsymbol{h}^{(s-1)}} = \boldsymbol{W} ^T \cdot diag(g'(\boldsymbol{b} + \boldsymbol{W}\cdot \boldsymbol{h}^{(s-1)} + \boldsymbol{U}\cdot \boldsymbol{x}^{(s)})) = \boldsymbol{W} ^T \cdot diag(g'(\boldsymbol{a}^{(s)})).

* If you see the figure as a type of graphical model, you should be able to understand the why chain rules can be applied as the equation above.

*According to this paper \frac{\partial \boldsymbol{h}^{(s)}}{\partial \boldsymbol{h}^{(s-1)}}  = \boldsymbol{W} ^T \cdot diag(g'(\boldsymbol{a}^{(s)})), but it seems that many study materials and web sites are mistaken in this point.

Hence \frac{\partial J^{(t)}}{\partial \boldsymbol{h}^{(k)}} = \frac{\partial J^{(t)}}{\partial \boldsymbol{h}^{(t)}} \prod_{k< s \leq t} \boldsymbol{W} ^T \cdot diag(g'(\boldsymbol{a}^{(s)})) = \frac{\partial J^{(t)}}{\partial \boldsymbol{h}^{(t)}} (\boldsymbol{W} ^T )^{(t - k)} \prod_{k< s \leq t} diag(g'(\boldsymbol{a}^{(s)})). If you take norms of \frac{\partial J^{(t)}}{\partial \boldsymbol{h}^{(k)}} you get an equality \left\lVert \frac{\partial J^{(t)}}{\partial \boldsymbol{h}^{(k)}} \right\rVert \leq \left\lVert \frac{\partial J^{(t)}}{\partial \boldsymbol{h}^{(t)}} \right\rVert \left\lVert \boldsymbol{W} ^T \right\rVert ^{(t - k)} \prod_{k< s \leq t} \left\lVert diag(g'(\boldsymbol{a}^{(s)}))\right\rVert. I will not go into detail anymore, but it is known that according to this inequality, multiplication of weight vectors exponentially converge to 0 or to infinite number.

We have seen that the error \frac{\partial J^{(t)}}{\partial \boldsymbol{h}^{(k)}} is the main factor causing vanishing/exploding gradient problems of simple RNNs. In case of LSTM, \frac{\partial J^{(t)}}{\partial \boldsymbol{c}^{(k)}} is an equivalent. For simplicity, let’s calculate only \frac{\partial \boldsymbol{c}^{(t)}}{\partial \boldsymbol{c}^{(t-1)}}, which is equivalent to \frac{\partial \boldsymbol{h}^{(t)}}{\partial \boldsymbol{h}^{(t-1)}} of simple RNN backprop.

* Just as I noted above, you have to be careful of which part the partial differential operator \frac{\partial}{\partial \boldsymbol{c}^{(t-1)}} affects in the chain rule above. That is, you need to calculate \frac{\partial}{\partial \boldsymbol{c}^{(t-1)}} (\boldsymbol{c}^{(t-1)} \odot \boldsymbol{f}^{(t)}), and the partial differential operator only affects \boldsymbol{c}^{(t-1)}. I think this is not a correct mathematical notation, but please forgive me for doing this for convenience.

If you continue calculating the equation above more concretely, you get the equation below.

I cannot mathematically explain why, but it is known that this characteristic of gradients of LSTM backprop mitigate the vanishing/exploding gradient problem. We have seen that if you take a norm of \frac{\partial J^{(t)}}{\partial \boldsymbol{h}^{(k)}}, that is equal or smaller than repeated multiplication of the norm of the same weight matrix, and that soon leads to vanishing/exploding gradient problem. But according to the equation above, even if you take a norm of repeatedly multiplied \frac{\partial \boldsymbol{c}^{(t)}}{\partial \boldsymbol{c}^{(t-1)}}, its norm cannot be evaluated with a simple value like repeated multiplication of the norm of the same weight matrix. The outputs of each gate are different from time steps to time steps, and that adjust the value of \frac{\partial \boldsymbol{c}^{(t)}}{\partial \boldsymbol{c}^{(t-1)}}.

*I personally guess the term diag(\boldsymbol{f}^{(t)}) is very effective. The unaffected value of the elements of \boldsymbol{f}^{(t)} can directly adjust the value of \frac{\partial \boldsymbol{c}^{(t)}}{\partial \boldsymbol{c}^{(t-1)}}. And as a matter of fact, it is known that performances of LSTM drop the most when you get rid of forget gates.

When it comes to tackling exploding gradient problems, there is a much easier technique called gradient clipping. This algorithm is very simple: you just have to adjust the size of gradient so that the absolute value of gradient is under a threshold at every time step. Imagine that you decide in which direction to move by calculating gradients, but when the footstep is going to be too big, you just adjust the size of footstep to the threshold size you have set. In a pseudo code, you can write a gradient clipping part only with some two line codes as below.

*\boldsymbol{g} is a gradient at the time step threshold is the maximum size of the “step.”

The figure below, cited from a deep learning text from MIT press textbook, is a good and straightforward explanation on gradient clipping.It is known that a strongly nonlinear function, such as error functions of RNN, can have very steep or plain areas. If you artificially visualize the idea in 3-dimensional space, as the surface of a loss function J with two variants w, b, that means the loss function J has plain areas and very steep cliffs like in the figure.Without gradient clipping, at the left side, you can see that the black dot all of a sudden climb the cliff and could jump to an unexpected area. But with gradient clipping, you avoid such “big jumps” on error functions.

Source: Source: Goodfellow and Yoshua Bengio and Aaron Courville, Deep Learning, (2016), MIT Press, 409p


I am glad that I have finally finished this article series. I am not sure how many of the readers would have read through all of the articles in this series, including my PowerPoint slides. I bet that is not so many. I spent a great deal of my time for making these contents, but sadly even when I was studying LSTM, it was becoming old-fashioned, at least in natural language processing (NLP) field: a very promising algorithm named Transformer has been replacing the position of LSTM. Deep learning is a very fast changing field. I also would like to make illustrative introductions on attention mechanism in NLP, from seq2seq model to Transformer. And I think LSTM would still remain as one of the algorithms in sequence data processing, such as hidden Hidden Markov model, or particle filter.


* I make study materials on machine learning, sponsored by DATANOMIQ. I do my best to make my content as straightforward but as precise as possible. I include all of my reference sources. If you notice any mistakes in my materials, including grammatical errors, please let me know (email: And if you have any advice for making my materials more understandable to learners, I would appreciate hearing it.

Simple RNN

Understanding LSTM forward propagation in two ways

*This article is only for the sake of understanding the equations in the second page of the paper named “LSTM: A Search Space Odyssey”. If you have no trouble understanding the equations of LSTM forward propagation, I recommend you to skip this article and go the the next article.

*This article is the fourth article of “A gentle introduction to the tiresome part of understanding RNN.”

1. Preface

I  heard that in Western culture, smart people write textbooks so that other normal people can understand difficult stuff, and that is why textbooks in Western countries tend to be bulky, but also they are not so difficult as they look. On the other hand in Asian culture, smart people write puzzling texts on esoteric topics, and normal people have to struggle to understand what noble people wanted to say. Publishers also require the authors to keep the texts as short as possible, so even though the textbooks are thin, usually students have to repeat reading the textbooks several times because usually they are too abstract.

Both styles have cons and pros, and usually I prefer Japanese textbooks because they are concise, and sometimes it is annoying to read Western style long texts with concrete straightforward examples to reach one conclusion. But a problem is that when it comes to explaining LSTM, almost all the text books are like Asian style ones. Every study material seems to skip the proper steps necessary for “normal people” to understand its algorithms. But after actually making concrete slides on mathematics on LSTM, I understood why: if you write down all the equations on LSTM forward/back propagation, that is going to be massive, and actually I had to make 100-page PowerPoint animated slides to make it understandable to people like me.

I already had a feeling that “Does it help to understand only LSTM with this precision? I should do more practical codings.” For example François Chollet, the developer of Keras, in his book, said as below.


For me that sounds like “We have already implemented RNNs for you, so just shut up and use Tensorflow/Keras.” Indeed, I have never cared about the architecture of my Mac Book Air, but I just use it every day, so I think he is to the point. To make matters worse, for me, a promising algorithm called Transformer seems to be replacing the position of LSTM in natural language processing. But in this article series and in my PowerPoint slides, I tried to explain as much as possible, contrary to his advice.

But I think, or rather hope,  it is still meaningful to understand this 23-year-old algorithm, which is as old as me. I think LSTM did build a generation of algorithms for sequence data, and actually Sepp Hochreiter, the inventor of LSTM, has received Neural Network Pioneer Award 2021 for his work.

I hope those who study sequence data processing in the future would come to this article series, and study basics of RNN just as I also study classical machine learning algorithms.

 *In this article “Densely Connected Layers” is written as “DCL,” and “Convolutional Neural Network” as “CNN.”

2. Why LSTM?

First of all, let’s take a brief look at what I said about the structures of RNNs,  in the first and the second article. A simple RNN is basically densely connected network with a few layers. But the RNN gets an input every time step, and it gives out an output at the time step. Part of information in the middle layer are succeeded to the next time step, and in the next time step, the RNN also gets an input and gives out an output. Therefore, virtually a simple RNN behaves almost the same way as densely connected layers with many layers during forward/back propagation if you focus on its recurrent connections.

That is why simple RNNs suffer from vanishing/exploding gradient problems, where the information exponentially vanishes or explodes when its gradients are multiplied many times through many layers during back propagation. To be exact, I think you need to consider this problem precisely like you can see in this paper. But for now, please at least keep it in mind that when you calculate a gradient of an error function with respect to parameters of simple neural networks, you have to multiply parameters many times like below, and this type of calculation usually leads to vanishing/exploding gradient problem.

LSTM was invented as a way to tackle such problems as I mentioned in the last article.

3. How to display LSTM

I would like you to just go to image search on Google, Bing, or Yahoo!, and type in “LSTM.” I think you will find many figures, but basically LSTM charts are roughly classified into two types: in this article I call them “Space Odyssey type” and “electronic circuit type”, and in conclusion, I highly recommend you to understand LSTM as the “electronic circuit type.”

*I just randomly came up with the terms “Space Odyssey type” and “electronic circuit type” because the former one is used in the paper I mentioned, and the latter one looks like an electronic circuit to me. You do not have to take how I call them seriously.

However, not that all the well-made explanations on LSTM use the “electronic circuit type,” and I am sure you sometimes have to understand LSTM as the “space odyssey type.” And the paper “LSTM: A Search Space Odyssey,” which I learned a lot about LSTM from,  also adopts the “Space Odyssey type.”

LSTM architectur visualization

The main reason why I recommend the “electronic circuit type” is that its behaviors look closer to that of simple RNNs, which you would have seen if you read my former articles.

*Behaviors of both of them look different, but of course they are doing the same things.

If you have some understanding on DCL, I think it was not so hard to understand how simple RNNs work because simple RNNs  are mainly composed of linear connections of neurons and weights, whose structures are the same almost everywhere. And basically they had only straightforward linear connections as you can see below.

But from now on, I would like you to give up the ideas that LSTM is composed of connections of neurons like the head image of this article series. If you do that, I think that would be chaotic and I do not want to make a figure of it on Power Point. In short, sooner or later you have to understand equations of LSTM.

4. Forward propagation of LSTM in “electronic circuit type”

*For further understanding of mathematics of LSTM forward/back propagation, I recommend you to download my slides.

The behaviors of an LSTM block is quite similar to that of a simple RNN block: an RNN block gets an input every time step and gets information from the RNN block of the last time step, via recurrent connections. And the block succeeds information to the next block.

Let’s look at the simplified architecture of  an LSTM block. First of all, you should keep it in mind that LSTM have two streams of information: the one going through all the gates, and the one going through cell connections, the “highway” of LSTM block. For simplicity, we will see the architecture of an LSTM block without peephole connections, the lines in blue. The flow of information through cell connections is relatively uninterrupted. This helps LSTMs to retain information for a long time.

In a LSTM block, the input and the output of the former time step separately go through sections named “gates”: input gate, forget gate, output gate, and block input. The outputs of the forget gate, the input gate, and the block input join the highway of cell connections to renew the value of the cell.

*The small two dots on the cell connections are the “on-ramp” of cell conection highway.

*You would see the terms “input gate,” “forget gate,” “output gate” almost everywhere, but how to call the “block gate” depends on textbooks.

Let’s look at the structure of an LSTM block a bit more concretely. An LSTM block at the time step (t) gets \boldsymbol{y}^{(t-1)}, the output at the last time step,  and \boldsymbol{c}^{(t-1)}, the information of the cell at the time step (t-1), via recurrent connections. The block at time step (t) gets the input \boldsymbol{x}^{(t)}, and it separately goes through each gate, together with \boldsymbol{y}^{(t-1)}. After some calculations and activation, each gate gives out an output. The outputs of the forget gate, the input gate, the block input, and the output gate are respectively \boldsymbol{f}^{(t)}, \boldsymbol{i}^{(t)}, \boldsymbol{z}^{(t)}, \boldsymbol{o}^{(t)}. The outputs of the gates are mixed with \boldsymbol{c}^{(t-1)} and the LSTM block gives out an output \boldsymbol{y}^{(t)}, and gives \boldsymbol{y}^{(t)} and \boldsymbol{c}^{(t)} to the next LSTM block via recurrent connections.

You calculate \boldsymbol{f}^{(t)}, \boldsymbol{i}^{(t)}, \boldsymbol{z}^{(t)}, \boldsymbol{o}^{(t)} as below.

  • \boldsymbol{f}^{(t)}= \sigma(\boldsymbol{W}_{for} \boldsymbol{x}^{(t)} + \boldsymbol{R}_{for} \boldsymbol{y}^{(t-1)} +  \boldsymbol{b}_{for})
  • \boldsymbol{i}^{(t)}=\sigma(\boldsymbol{W}_{in} \boldsymbol{x}^{(t)} + \boldsymbol{R}_{in} \boldsymbol{y}^{(t-1)} + \boldsymbol{b}_{in})
  • \boldsymbol{z}^{(t)}=tanh(\boldsymbol{W}_z \boldsymbol{x}^{(t)} + \boldsymbol{R}_z \boldsymbol{y}^{(t-1)} + \boldsymbol{b}_z)
  • \boldsymbol{o}^{(t)}=\sigma(\boldsymbol{W}_{out} \boldsymbol{x}^{(t)} + \boldsymbol{R}_{out} \boldsymbol{y}^{(t-1)} + \boldsymbol{b}_{out})

*You have to keep it in mind that the equations above do not include peephole connections, which I am going to show with blue lines in the end.

The equations above are quite straightforward if you understand forward propagation of simple neural networks. You add linear products of \boldsymbol{y}^{(t)} and \boldsymbol{c}^{(t)} with different weights in each gate. What makes LSTMs different from simple RNNs is how to mix the outputs of the gates with the cell connections. In order to explain that, I need to introduce a mathematical operator called Hadamard product, which you denote as \odot. This is a very simple operator. This operator produces an elementwise product of two vectors or matrices with identical shape.

With this Hadamar product operator, the renewed cell and the output are calculated as below.

  • \boldsymbol{c}^{(t)} = \boldsymbol{z}^{(t)}\odot \boldsymbol{i}^{(t)} + \boldsymbol{c}^{(t-1)} \odot \boldsymbol{f}^{(t)}
  • \boldsymbol{y}^{(t)} = \boldsymbol{o}^{(t)} \odot tanh(\boldsymbol{c}^{(t)})

The values of \boldsymbol{f}^{(t)}, \boldsymbol{i}^{(t)}, \boldsymbol{z}^{(t)}, \boldsymbol{o}^{(t)} are compressed into the range of [0, 1] or [-1, 1] with activation functions. You can see that the input gate and the block input give new information to the cell. The part \boldsymbol{c}^{(t-1)} \odot \boldsymbol{f}^{(t)} means that the output of the forget gate “forgets” the cell of the last time step by multiplying the values from 0 to 1 elementwise. And the cell \boldsymbol{c}^{(t)} is activated with tanh() and the output of the output gate “suppress” the activated value of \boldsymbol{c}^{(t)}. In other words, the output gatedecides how much information to give out as an output of the LSTM block. The output of every gate depends on the input \boldsymbol{x}^{(t)}, and the recurrent connection \boldsymbol{y}^{(t-1)}. That means an LSTM block learns to forget the cell of the last time step, to renew the cell, and to suppress the output. To describe in an extreme manner, if all the outputs of every gate are always (1, 1, …1)^T, LSTMs forget nothing, retain information of inputs at every time step, and gives out everything. And  if all the outputs of every gate are always (0, 0, …0)^T, LSTMs forget everything, receive no inputs, and give out nothing.

This model has one problem: the outputs of each gate do not directly depend on the information in the cell. To solve this problem, some LSTM models introduce some flows of information from the cell to each gate, which are shown as lines in blue in the figure below.

LSTM inner architecture

LSTM models, for example the one with or without peephole connection, depend on the library you use, and the model I have showed is one of standard LSTM structure. However no matter how complicated structure of an LSTM block looks, you usually cover it with a black box as below and show its behavior in a very simplified way.

5. Space Odyssey type

I personally think there is no advantages of understanding how LSTMs work with this Space Odyssey type chart, but in several cases you would have to use this type of chart. So I will briefly explain how to look at that type of chart, based on understandings of LSTMs you have gained through this article.

In Space Odyssey type of LSTM chart, at the center is a cell. Electronic circuit type of chart, which shows the flow of information of the cell as an uninterrupted “highway” in an LSTM block. On the other hand, in a Spacey Odyssey type of chart, the information of the cell rotate at the center. And each gate gets the information of the cell through peephole connections,  \boldsymbol{x}^{(t)}, the input at the time step (t) , sand \boldsymbol{y}^{(t-1)}, the output at the last time step (t-1), which came through recurrent connections. In Space Odyssey type of chart, you can more clearly see that the information of the cell go to each gate through the peephole connections in blue. Each gate calculates its output.

Just as the charts you have seen, the dotted line denote the information from the past. First, the information of the cell at the time step (t-1) goes to the forget gate and get mixed with the output of the forget cell In this process the cell is partly “forgotten.” Next, the input gate and the block input are mixed to generate part of new value of the the cell at time step  (t). And the partly “forgotten” \boldsymbol{c}^{(t-1)} goes back to the center of the block and it is mixed with the output of the input gate and the block input. That is how \boldsymbol{c}^{(t)} is renewed. And the value of new cell flow to the top of the chart, being mixed with the output of the output gate. Or you can also say the information of new cell is “suppressed” with the output gate.

I have finished the first four articles of this article series, and finally I am gong to write about back propagation of LSTM in the next article. I have to say what I have written so far is all for the next article, and my long long Power Point slides.


* I make study materials on machine learning, sponsored by DATANOMIQ. I do my best to make my content as straightforward but as precise as possible. I include all of my reference sources. If you notice any mistakes in my materials, including grammatical errors, please let me know (email: And if you have any advice for making my materials more understandable to learners, I would appreciate hearing it.


[1] Klaus Greff, Rupesh Kumar Srivastava, Jan Koutník, Bas R. Steunebrink, Jürgen Schmidhuber, “LSTM: A Search Space Odyssey,” (2017)

[2] Francois Chollet, Deep Learning with Python,(2018), Manning , pp. 202-204

[3] “Sepp Hochreiter receives IEEE CIS Neural Networks Pioneer Award 2021”, Institute of advanced research in artificial intelligence, (2020)

[4] Oketani Takayuki, “Machine Learning Professional Series: Deep Learning,” (2015), pp. 120-125
岡谷貴之 著, 「機械学習プロフェッショナルシリーズ 深層学習」, (2015), pp. 120-125

[5] Harada Tatsuya, “Machine Learning Professional Series: Image Recognition,” (2017), pp. 252-257
原田達也 著, 「機械学習プロフェッショナルシリーズ 画像認識」, (2017), pp. 252-257

[6] “Understandable LSTM ~ With the Current Trends,” Qiita, (2015)
「わかるLSTM ~ 最近の動向と共に」, Qiita, (2015)

Simple RNN

A brief history of neural nets: everything you should know before learning LSTM

This series is not a college course or something on deep learning with strict deadlines for assignments, so let’s take a detour from practical stuff and take a brief look at the history of neural networks.

The history of neural networks is also a big topic, which could be so long that I had to prepare another article series. And usually I am supposed to begin such articles with something like “The term ‘AI’ was first used by John McCarthy in Dartmouth conference 1956…” but you can find many of such texts written by people with much more experiences in this field. Therefore I am going to write this article from my point of view, as an intern writing articles on RNN, as a movie buff, and as one of many Japanese men who spent a great deal of childhood with video games.

We are now in the third AI boom, and some researchers say this boom began in 2006. A professor in my university said there we are now in a kind of bubble economy in machine learning/data science industry, but people used to say “Stop daydreaming” to AI researchers. The second AI winter is partly due to vanishing/exploding gradient problem of deep learning. And LSTM was invented as one way to tackle such problems, in 1997.

1, First AI boom

In the first AI boom, I think people were literally “daydreaming.” Even though the applications of machine learning algorithms were limited to simple tasks like playing chess, checker, or searching route of 2d mazes, and sometimes this time is called GOFAI (Good Old Fashioned AI).


Even today when someone use the term “AI” merely for tasks with neural networks, that amuses me because for me deep learning is just statistically and automatically training neural networks, which are capable of universal approximation, into some classifiers/regressors. Actually the algorithms behind that is quite impressive, but the structure of human brains is much more complicated. The hype of “AI” already started in this first AI boom. Let me take an example of machine translation in this video. In fact the research of machine translation already started in the early 1950s, and of  specific interest in the time was translation between English and Russian due to Cold War. In the first article of this series, I said one of the most famous applications of RNN is machine translation, such as Google Translation, DeepL. They are a type of machine translation called neural machine translation because they use neural networks, especially RNNs. Neural machine translation was an astonishing breakthrough around 2014 in machine translation field. The former major type of machine translation was statistical machine translation, based on statistical language models. And the machine translator in the first AI boom was rule base machine translators, which are more primitive than statistical ones.


The most remarkable invention in this time was of course perceptron by Frank Rosenblatt. Some people say that this is the first neural network. Even though you can implement perceptron with a-few-line codes in Python, obviously they did not have Jupyter Notebook in those days. The perceptron was implemented as a huge instrument named Mark 1 Perceptron, and it was composed of randomly connected wires. I do not precisely know how it works, but it was a huge effort to implement even the most primitive type of neural networks. They needed to use a big lighting fixture to get a 20*20 pixel image using 20*20 array of cadmium sulphide photocells. The research by Rosenblatt, however, was criticized by Marvin Minsky in his book because perceptrons could only be used for linearly separable data. To make matters worse the criticism prevailed as that more general, multi-layer perceptrons were also not useful for linearly inseparable data (as I mentioned in the first article, multi-layer perceptrons, namely normal neural networks,  can be universal approximators, which have potentials to classify/regress various types of complex data). In case you do not know what “linearly separable” means, imagine that there are data plotted on a piece of paper. If an elementary school kid can draw a border line between two clusters of the data with a ruler and a pencil on the paper, the 2d data is “linearly separable”….

With big disappointments to the research on “electronic brains,” the budget of AI research was reduced and AI research entered its first winter.

Source: and

I think  the frame problem (1969),  by John McCarthy and Patrick J. Hayes, is also an iconic theory in the end of the first AI boom. This theory is known as a story of creating a robot trying to pull out its battery on a wheeled wagon in a room. But there is also a time bomb on the wagon. The first prototype of the robot, named R1, naively tried to pull out the wagon form the room, and the bomb exploded. The problems was obvious: R1 was not programmed to consider the risks by taking each action, so the researchers made the next prototype named R1D1, which was programmed to consider the potential risks of taking each action. When R1D1 tried to pull out the wagon, it realized the risk of pulling the bomb together with the battery. But soon it started considering all the potential risks, such as the risk of the ceiling falling down, the distance between the wagon and all the walls, and so on, when the bomb exploded. The next problem was also obvious: R1D1 was not programmed to distinguish if the factors are relevant of irrelevant to the main purpose, and the next prototype R2D1 was programmed to do distinguish them. This time, R2D1 started thinking about “whether the factor is  irrelevant to the main purpose,” on every factor measured, and again the bomb exploded. How can we get a perfect AI, R2D2?

The situation of mentioned above is a bit extreme, but it is said AI could also get stuck when it try to take some super simple actions like finding a number in a phone book and make a phone call. It is difficult for an artificial intelligence to decide what is relevant and what is irrelevant, but humans will not get stuck with such simple stuff, and sometimes the frame problem is counted as the most difficult and essential problem of developing AI. But personally I think the original frame problem was unreasonable in that McCarthy, in his attempts to model the real world, was inflexible in his handling of the various equations involved, treating them all with equal weight regardless of the particular circumstances of a situation. Some people say that McCarthy, who was an advocate for AI, also wanted to see the field come to an end, due to its failure to meet the high expectations it once aroused.

Not only the frame problem, but also many other AI-related technological/philosophical problems have been proposed, such as Chinese room (1980), the symbol grounding problem (1990), and they are thought to be as hardships in inventing artificial intelligence, but I omit those topics in this article.

*The name R2D2 did not come from the famous story of frame problem. The story was Daniel Dennett first proposed the story of R2D2 in his paper published in 1984. Star Wars was first released in 1977. It is said that the name R2D2 came from “Reel 2, Dialogue 2,” which George Lucas said while film shooting. And the design of C3PO came from Maria in Metropolis(1927). It is said that the most famous AI duo in movie history was inspired by Tahei and Matashichi in The Hidden Fortress (1958), directed by Kurosawa Akira.


Interestingly, in the end of the first AI boom, 2001: A Space Odyssey, directed by Stanley Kubrick, was released in 1968. Unlike conventional fantasylike AI characters, for example Maria in Metropolis (1927), HAL 9000 was portrayed as a very realistic AI, and the movie already pointed out the risk of AI being insane when it gets some commands from several users. HAL 9000 still has been a very iconic character in AI field. For example when you say some quotes from 2001: A Space Odyssey to Siri you get some parody responses. I also thin you should keep it in mind that in order to make an AI like HAL 9000 come true, for now RNNs would be indispensable in many ways: you would need RNNs for better voice recognition, better conversational system, and for reading lips.


*Just as you cannot understand Monty Python references in Python official tutorials without watching Monty Python and the Holy Grail, you cannot understand many parodies in AI contexts without watching 2001: A Space Odyssey. Even though the movie had some interview videos with some researchers and some narrations, Stanley Kubrick cut off all the footage and made the movie very difficult to understand. Most people did not or do not understand that it is a movie about aliens who gave homework of coming to Jupiter to human beings.

2, Second AI boom/winter

Source: Fukushima Kunihiko, “Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position,” (1980)

I am not going to write about the second AI boom in detail, but at least you should keep it in mind that convolutional neural network (CNN) is a keyword in this time. Neocognitron, an artificial model of how sight nerves perceive thing, was invented by Kunihiko Fukushima in 1980, and the model is said to be the origin on CNN. And Neocognitron got inspired by the Hubel and Wiesel’s research on sight nerves. In 1989, a group in AT & T Bell Laboratory led by Yann LeCun invented the first practical CNN to read handwritten digit.

Y. LeCun, “Backpropagation Applied to Handwritten Zip Code Recognition,” (1989)

Another turning point in this second AI boom was that back propagation algorithm was discovered, and the CNN by LeCun was also trained with back propagation. LeCun made a deep neural networks with some layers in 1998 for more practical uses.

But his research did not gain so much attention like today, because AI research entered its second winter at the beginning of the 1990s, and that was partly due to vanishing/exploding gradient problem of deep learning. People knew that neural networks had potentials of universal approximation, but when they tried to train naively stacked neural nets, the gradients, which you need for training neural networks, exponentially increased/decreased. Even though the CNN made by LeCun was the first successful case of “deep” neural nets which did not suffer from the vanishing/exploding gradient problem so much, deep learning research also stagnated in this time.

The ultimate goal of this article series is to understand LSTM at a more abstract/mathematical level because it is one of the practical RNNs, but the idea of LSTM (Long Short Term Memory) itself was already proposed in 1997 as an RNN algorithm to tackle vanishing gradient problem. (Exploding gradient problem is solved with a technique named gradient clipping, and this is easier than techniques for preventing vanishing gradient problems. I am also going to explain it in the next article.) After that some other techniques like introducing forget gate, peephole connections, were discovered, but basically it took some 20 years till LSTM got attentions like today. The reasons for that is lack of hardware and data sets, and that was also major reasons for the second AI winter.

Source: Sepp HochreiterJürgen, Schmidhuber, “Long Short-term Memory,” (1997)

In the 1990s, the mid of second AI winter, the Internet started prevailing for commercial uses. I think one of the iconic events in this time was the source codes WWW (World Wide Web) were announced in 1993. Some of you might still remember that you little by little became able to transmit more data online in this time. That means people came to get more and more access to various datasets in those days, which is indispensable for machine learning tasks.

After all, we could not get HAL 9000 by the end of 2001, but instead we got Xbox console.

3, Video game industry and GPU

Even though research on neural networks stagnated in the 1990s the same period witnessed an advance in the computation of massive parallel linear transformations, due to their need in fields such as image processing.

Computer graphics move or rotate in 3d spaces, and that is also linear transformations. When you think about a car moving in a city, it is convenient to place the car, buildings, and other objects on a fixed 3d space. But when you need to make computer graphics of scenes of the city from a view point inside the car, you put a moving origin point in the car and see the city. The spatial information of the city is calculated as vectors from the moving origin point. Of course this is also linear transformations. Of course I am not talking about a dot or simple figures moving in the 3d spaces. Computer graphics are composed of numerous plane panels, and each of them have at least three vertexes, and they move on 3d spaces. Depending on viewpoints, you need project the 3d graphics in 3d spaces on 2d spaces to display the graphics on devices. You need to calculate which part of the panel is projected to which pixel on the display, and that is called rasterization. Plus, in order to get photophotorealistic image, you need to think about how lights from light sources reflect on the panel and projected on the display. And you also have to put some textures on groups of panels. You might also need to change color spaces, which is also linear transformations.

My point is, in short, you really need to do numerous linear transformations in parallel in image processing.

When it comes to the use of CGI in movies,  two pioneer movies were released during this time: Jurassic Park in 1993, and Toy Story in 1995. It is famous that Pixar used to be one of the departments in ILM (Industrial Light and Magic), founded by George Lucas, and Steve Jobs bought the department. Even though the members in Pixar had not even made a long feature film in their lives, after trial and errors, they made the first CGI animated feature movie. On the other hand, in order to acquire funds for the production of Schindler’s List (1993), Steven Spielberg took on Jurassic Park (1993), consequently changing the history of CGI through this “side job.”


*I think you have realized that George Lucas is mentioned almost everywhere in this article. His influences on technologies are not only limited to image processing, but also sound measuring system, nonlinear editing system. Photoshop was also originally developed under his company. I need another article series for this topic, but maybe not in Data Science Blog.


Considering that the first wire-frame computer graphics made and displayed by computers appeared in the scene of displaying the wire frame structure of Death Star in a war room, in Star Wars: A New Hope, the development of CGI was already astonishing at this time. But I think deep learning owe its development more to video game industry.

*I said that the Death Star scene is the first use of graphics made and DISPLAYED by computers, because I have to say one of the first graphics in movie MADE by computer dates back to the legendary title sequence of Vertigo(1958).

When it comes to 3D video games the processing unit has to constantly deal with real time commands from controllers. It is famous that GPU was originally specifically designed for plotting computer graphics. Video game market is the biggest in entertainment industry in general, and it is said that the quality of computer graphics have the strongest correlation with video games sales, therefore enhancing this quality is a priority for the video game console manufacturers.

One good example to see how much video games developed is comparing original Final Fantasy 7 and the remake one. The original one was released in 1997, the same year as when LSTM was invented. And recently  the remake version of Final Fantasy 7 was finally released this year. The original one was also made with very big budget, and it was divided into three CD-ROMs. The original one was also very revolutionary given that the former ones of Final Fantasy franchise were all 2d video retro style video games. But still the computer graphics looks like polygons, and in almost all scenes the camera angle was fixed in the original one. On the other hand the remake one is very photorealistic and you can move the angle of the camera as you want while you play the video game.

There were also fierce battles by graphic processor manufacturers in computer video game market in the 1990s, but personally I think the release of Xbox console was a turning point in the development of GPU. To be concrete, Microsoft adopted a type of NV20 GPU for Xbox consoles, and that left some room of programmability for developers. The chief architect of NV20, which was released under the brand of GeForce3, said making major changes in the company’s graphic chips was very risky. But that decision opened up possibilities of uses of GPU beyond computer graphics.


I think that the idea of a programmable GPU provided other scientific fields with more visible benefits after CUDA was launched. And GPU gained its position not only in deep learning, but also many other fields including making super computers.

*When it comes to deep learning, even GPUs have strong rivals. TPU(Tensor Processing Unit) made by Google, is specialized for deep learning tasks, and have astonishing processing speed. And FPGA(Field Programmable Gate Array), which was originally invented customizable electronic circuit, proved to be efficient for reducing electricity consumption of deep learning tasks.

*I am not so sure about this GPU part. Processing unit, including GPU is another big topic, that is beyond my capacity to be honest.  I would appreciate it if you could share your view and some references to confirm your opinion, on the comment section or via email.

*If you are interested you should see this video of game fans’ reactions to the announcement of Final Fantasy 7. This is the industry which grew behind the development of deep learning, and many fields where you need parallel computations owe themselves to the nerds who spent a lot of money for video games, including me.

*But ironically the engineers who invented the GPU said they did not play video games simply because they were busy. If you try to study the technologies behind video games, you would not have much time playing them. That is the reality.

We have seen that the in this second AI winter, Internet and GPU laid foundation of the next AI boom. But still the last piece of the puzzle is missing: let’s look at the breakthrough which solved the vanishing /exploding gradient problem of deep learning in the next section.

4, Pretraining of deep belief networks: “The Dawn of Deep Learning”

Some researchers say the invention of pretraining of deep belief network by Geoffrey Hinton was a breakthrough which put an end to the last AI winter. Deep belief networks are different type of networks from the neural networks we have discussed, but their architectures are similar to those of the neural networks. And it was also unknown how to train deep belief nets when they have several layers. Hinton discovered that training the networks layer by layer in advance can tackle vanishing gradient problems. And later it was discovered that you can do pretraining neural networks layer by layer with autoencoders.

*Deep belief network is beyond the scope of this article series. I have to talk about generative models, Boltzmann machine, and some other topics.

The pretraining techniques of neural networks is not mainstream anymore. But I think it is very meaningful to know that major deep learning techniques such as using ReLU activation functions, optimization with Adam, dropout, batch normalization, came up as more effective algorithms for deep learning after the advent of the pretraining techniques, and now we are in the third AI boom.

In the next next article we are finally going to work on LSTM. Specifically, I am going to offer a clearer guide to a well-made paper on LSTM, named “LSTM: A Search Space Odyssey.”

* I make study materials on machine learning, sponsored by DATANOMIQ. I do my best to make my content as straightforward but as precise as possible. I include all of my reference sources. If you notice any mistakes in my materials, including grammatical errors, please let me know (email: And if you have any advice for making my materials more understandable to learners, I would appreciate hearing it.


[1] Taniguchi Tadahiro, “An Illustrated Guide to Artificial Intelligence”, (2010), Kodansha pp. 3-11
谷口忠大 著, 「イラストで学ぶ人工知能概論」, (2010), 講談社, pp. 3-11

[2] Francois Chollet, Deep Learning with Python,(2018), Manning , pp. 14-24

[3] Oketani Takayuki, “Machine Learning Professional Series: Deep Learning,” (2015), pp. 1-5, 151-156
岡谷貴之 著, 「機械学習プロフェッショナルシリーズ 深層学習」, (2015), pp. 1-5, 151-156

[4] Abigail See, Matthew Lamm, “Natural Language Processingwith Deep LearningCS224N/Ling284 Lecture 8:Machine Translation,Sequence-to-sequence and Attention,” (2020),

[5]C. M. Bishop, “Pattern Recognition and Machine Learning,” (2006), Springer, pp. 192-196

[6] Daniel C. Dennett, “Cognitive Wheels: the Frame Problem of AI,” (1984), pp. 1-2

[7] Machiyama Tomohiro, “Understanding Cinemas of 1967-1979,” (2014), Yosensya, pp. 14-30
町山智浩 著, 「<映画の見方>が分かる本」,(2014), 洋泉社, pp. 14-30

[8] Harada Tatsuya, “Machine Learning Professional Series: Image Recognition,” (2017), pp. 156-157
原田達也 著, 「機械学習プロフェッショナルシリーズ 画像認識」, (2017), pp. 156-157

[9] Suyama Atsushi, “Machine Learning Professional Series: Bayesian Deep Learning,” (2019)岡谷貴之 須山敦志 著, 「機械学習プロフェッショナルシリーズ ベイズ深層学習」, (2019)

[10] “Understandable LSTM ~ With the Current Trends,” Qiita, (2015)
「わかるLSTM ~ 最近の動向と共に」, Qiita, (2015)

[11] Hisa Ando, “WEB+DB PRESS plus series: Technologies Supporting Processors – The World Endlessly Pursuing Speed,” (2017), Gijutsu-hyoron-sya, pp 313-317
Hisa Ando, 「WEB+DB PRESS plusシリーズ プロセッサを支える技術― 果てしなくスピードを追求する世界」, (2017), 技術評論社, pp. 313-317

[12] “Takahashi Yoshiki and Utamaru discuss George Lucas,” miyearnZZ Labo, (2016)
“高橋ヨシキと宇多丸 ジョージ・ルーカスを語る,” miyearnZZ Labo, (2016)

[13] Katherine Bourzac, “Chip Hall of Fame: Nvidia NV20 The first configurable graphics processor opened the door to a machine-learning revolution,” IEEE SPECTRUM, (2018)

Data Science für Smart Home im familiengeführten Unternehmen Miele

Dr. Florian Nielsen ist Principal for AI und Data Science bei Miele im Bereich Smart Home und zuständig für die Entwicklung daten-getriebener digitaler Produkte und Produkterweiterungen. Der studierte Informatiker promovierte an der Universität Ulm zum Thema multimodale kognitive technische Systeme.

Data Science Blog: Herr Dr. Nielsen, viele Unternehmen und Anwender reden heute schon von Smart Home, haben jedoch eher ein Remote Home. Wie machen Sie daraus tatsächlich ein Smart Home?

Tatsächlich entspricht das auch meiner Wahrnehmung. Die bloße Steuerung vernetzter Produkte über digitale Endgeräte macht aus einem vernetzten Produkt nicht gleich ein „smartes“. Allerdings ist diese Remotefunktion ein notwendiges Puzzlestück in der Entwicklung von einem nicht vernetzten Produkt, über ein intelligentes, vernetztes Produkt hin zu einem Ökosystem von sich ergänzenden smarten Produkten und Services. Vernetzte Produkte, selbst wenn sie nur aus der Ferne gesteuert werden können, erzeugen Daten und ermöglichen uns die Personalisierung, Optimierung oder gar Automatisierung von Produktfunktionen basierend auf diesen Daten voran zu treiben. „Smart“ wird für mich ein Produkt, wenn es sich beispielsweise besser den Bedürfnissen des Nutzers anpasst oder über Assistenzfunktionen eine Arbeitserleichterung im Alltag bietet.

Data Science Blog: Smart Home wiederum ist ein großer Begriff, der weit mehr als Geräte für Küchen und Badezimmer betrifft. Wie weit werden Sie hier ins Smart Home vordringen können?

Smart Home ist für mich schon fast ein verbrannter Begriff. Der Nutzer assoziiert hiermit doch vor allem die Steuerung von Heizung und Rollladen. Im Prinzip geht es doch um eine Vision in der sich smarte, vernetzte Produkt in ein kontextbasiertes Ökosystem einbetten um den jeweiligen Nutzer in seinem Alltag, nicht nur in seinem Zuhause, Mehrwert mit intelligenten Produkten und Services zu bieten. Für uns fängt das beispielsweise nicht erst beim Starten des Kochprozesses mit Miele-Geräten an, sondern deckt potenziell die komplette „User Journey“ rund um Ernährung (z. B. Inspiration, Einkaufen, Vorratshaltung) und Kochen ab. Natürlich überlegen wir verstärkt, wie Produkte und Services unser existierendes Produktportfolio ergänzen bzw. dem Nutzer zugänglicher machen könnten, beschränken uns aber hierauf nicht. Ein zusätzlicher für uns als Miele essenzieller Aspekt ist allerdings auch die Privatsphäre des Kunden. Bei der Bewertung potenzieller Use-Cases spielt die Privatsphäre unserer Kunden immer eine wichtige Rolle.

Data Science Blog: Die meisten Data-Science-Abteilungen befassen sich eher mit Prozessen, z. B. der Qualitätsüberwachung oder Prozessoptimierung in der Produktion. Sie jedoch nutzen Data Science als Komponente für Produkte. Was gibt es dabei zu beachten?

Kundenbedürfnisse. Wir glauben an nutzerorientierte Produktentwicklung und dementsprechend fängt alles bei uns bei der Identifikation von Bedürfnissen und potenziellen Lösungen hierfür an. Meist starten wir mit „Design Thinking“ um die Themen zu identifizieren, die für den Kunden einen echten Mehrwert bieten. Wenn dann noch Data Science Teil der abgeleiteten Lösung ist, kommen wir verstärkt ins Spiel. Eine wesentliche Herausforderung ist, dass wir oft nicht auf der grünen Wiese starten können. Zumindest wenn es um ein zusätzliches Produktfeature geht, das mit bestehender Gerätehardware, Vernetzungsarchitektur und der daraus resultierenden Datengrundlage zurechtkommen muss. Zwar sind unsere neuen Produktgenerationen „Remote Update“-fähig, aber auch das hilft uns manchmal nur bedingt. Dementsprechend ist die Antizipation von Geräteanforderungen essenziell. Etwas besser sieht es natürlich bei Umsetzungen von cloud-basierten Use-Cases aus.

Data Science Blog: Es heißt häufig, dass Data Scientists kaum zu finden sind. Ist Recruiting für Sie tatsächlich noch ein Thema?

Data Scientists, hier mal nicht interpretiert als Mythos „Unicorn“ oder „Full-Stack“ sind natürlich wichtig, und auch nicht leicht zu bekommen in einer Region wie Gütersloh. Aber Engineers, egal ob Data, ML, Cloud oder Software generell, sind der viel wesentlichere Baustein für uns. Für die Umsetzung von Ideen braucht es nun mal viel Engineering. Es ist mittlerweile hinlänglich bekannt, dass Data Science einen zwar sehr wichtigen, aber auch kleineren Teil des daten-getriebenen Produkts ausmacht. Mal abgesehen davon habe ich den Eindruck, dass immer mehr „Data Science“- Studiengänge aufgesetzt werden, die uns einerseits die Suche nach Personal erleichtern und andererseits ermöglichen Fachkräfte einzustellen die nicht, wie früher einen PhD haben (müssen).

Data Science Blog: Sie haben bereits einige Analysen erfolgreich in Ihre Produkte integriert. Welche Herausforderungen mussten dabei überwunden werden? Und welche haben Sie heute noch vor sich?

Wir sind, wie viele Data-Science-Abteilungen, noch ein relativ junger Bereich. Bei den meisten unserer smarten Produkte und Services stecken wir momentan in der MVP-Entwicklung, deshalb gibt es einige Herausforderungen, die wir aktuell hautnah erfahren. Dies fängt, wie oben erwähnt, bei der Berücksichtigung von bereits vorhandenen Gerätevoraussetzungen an, geht über mitunter heterogene, inkonsistente Datengrundlagen, bis hin zur Etablierung von Data-Science- Infrastruktur und Deploymentprozessen. Aus meiner Sicht stehen zudem viele Unternehmen vor der Herausforderung die Weiterentwicklung und den Betrieb von AI bzw. Data- Science- Produkten sicherzustellen. Verglichen mit einem „fire-and-forget“ Mindset nach Start der Serienproduktion früherer Zeiten muss ein Umdenken stattfinden. Daten-getriebene Produkte und Services „leben“ und müssen dementsprechend anders behandelt und umsorgt werden – mit mehr Aufwand aber auch mit der Chance „immer besser“ zu werden. Deshalb werden wir Buzzwords wie „MLOps“ vermehrt in den üblichen Beraterlektüren finden, wenn es um die nachhaltige Generierung von Mehrwert von AI und Data Science für Unternehmen geht. Und das zu Recht.

Data Science Blog: Data Driven Thinking wird heute sowohl von Mitarbeitern in den Fachbereichen als auch vom Management verlangt. Gerade für ein Traditionsunternehmen wie Miele sicherlich eine Herausforderung. Wie könnten Sie diese Denkweise im Unternehmen fördern?

Data Driven Thinking kann nur etabliert werden, wenn überhaupt der Zugriff auf Daten und darauf aufbauende Analysen gegeben ist. Deshalb ist Daten-Demokratisierung der wichtigste erste Schritt. Aus meiner Perspektive geht es darum initial die Potenziale aufzuzeigen, um dann mithilfe von Daten Unsicherheiten zu reduzieren. Wir haben die Erfahrung gemacht, dass viele Fachbereiche echtes Interesse an einer daten-getriebenen Analyse ihrer Hypothesen haben und dankbar für eine daten-getriebene Unterstützung sind. Miele war und ist ein sehr innovatives Unternehmen, dass „immer besser“ werden will. Deshalb erfahren wir momentan große Unterstützung von ganz oben und sind sehr positiv gestimmt. Wir denken, dass ein Schritt in die richtige Richtung bereits getan ist und mit zunehmender Zahl an Multiplikatoren ein „Data Driven Thinking“ sich im gesamten Unternehmen etablieren kann.