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.

  • On the difficulty of language: prerequisites for NLP with Transformer (Coming soon)
  • Seq2seq model and attention mechanism: a backbone of NLP with deep learning (Coming soon)
  • Multi-head attention: the key component of Transformer (Coming soon)
  • The whole architecture of Transformer and with my toy English/German translator (Coming soon)
  • Transformer in image processing (Coming soon)

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.

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.

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.

 

[References]

[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)
URL: https://www.iarai.ac.at/news/sepp-hochreiter-receives-ieee-cis-neural-networks-pioneer-award-2021/?fbclid=IwAR27cwT5MfCw4Tqzs3MX_W9eahYDcIFuoGymATDR1A-gbtVmDpb8ExfQ87A

[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)
URL: https://qiita.com/t_Signull/items/21b82be280b46f467d1b

Simple RNN

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

This 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.

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. 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 thought 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

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.

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 to train 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, 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.

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: yasuto.tamura@datanomiq.de). And if you have any advice for making my materials more understandable to learners, I would appreciate hearing it.

Data Analytics and Mining for Dummies

Data Analytics and Mining is often perceived as an extremely tricky task cut out for Data Analysts and Data Scientists having a thorough knowledge encompassing several different domains such as mathematics, statistics, computer algorithms and programming. However, there are several tools available today that make it possible for novice programmers or people with no absolutely no algorithmic or programming expertise to carry out Data Analytics and Mining. One such tool which is very powerful and provides a graphical user interface and an assembly of nodes for ETL: Extraction, Transformation, Loading, for modeling, data analysis and visualization without, or with only slight programming is the KNIME Analytics Platform.

KNIME, or the Konstanz Information Miner, was developed by the University of Konstanz and is now popular with a large international community of developers. Initially KNIME was originally made for commercial use but now it is available as an open source software and has been used extensively in pharmaceutical research since 2006 and also a powerful data mining tool for the financial data sector. It is also frequently used in the Business Intelligence (BI) sector.

KNIME as a Data Mining Tool

KNIME is also one of the most well-organized tools which enables various methods of machine learning and data mining to be integrated. It is very effective when we are pre-processing data i.e. extracting, transforming, and loading data.

KNIME has a number of good features like quick deployment and scaling efficiency. It employs an assembly of nodes to pre-process data for analytics and visualization. It is also used for discovering patterns among large volumes of data and transforming data into more polished/actionable information.

Some Features of KNIME:

  • Free and open source
  • Graphical and logically designed
  • Very rich in analytics capabilities
  • No limitations on data size, memory usage, or functionalities
  • Compatible with Windows ,OS and Linux
  • Written in Java and edited with Eclipse.

A node is the smallest design unit in KNIME and each node serves a dedicated task. KNIME contains graphical, drag-drop nodes that require no coding. Nodes are connected with one’s output being another’s input, as a workflow. Therefore end-to-end pipelines can be built requiring no coding effort. This makes KNIME stand out, makes it user-friendly and make it accessible for dummies not from a computer science background.

KNIME workflow designed for graduate admission prediction

KNIME workflow designed for graduate admission prediction

KNIME has nodes to carry out Univariate Statistics, Multivariate Statistics, Data Mining, Time Series Analysis, Image Processing, Web Analytics, Text Mining, Network Analysis and Social Media Analysis. The KNIME node repository has a node for every functionality you can possibly think of and need while building a data mining model. One can execute different algorithms such as clustering and classification on a dataset and visualize the results inside the framework itself. It is a framework capable of giving insights on data and the phenomenon that the data represent.

Some commonly used KNIME node groups include:

  • Input-Output or I/O:  Nodes in this group retrieve data from or to write data to external files or data bases.
  • Data Manipulation: Used for data pre-processing tasks. Contains nodes to filter, group, pivot, bin, normalize, aggregate, join, sample, partition, etc.
  • Views: This set of nodes permit users to inspect data and analysis results using multiple views. This gives a means for truly interactive exploration of a data set.
  • Data Mining: In this group, there are nodes that implement certain algorithms (like K-means clustering, Decision Trees, etc.)

Comparison with other tools 

The first version of the KNIME Analytics Platform was released in 2006 whereas Weka and R studio were released in 1997 and 1993 respectively. KNIME is a proper data mining tool whereas Weka and R studio are Machine Learning tools which can also do data mining. KNIME integrates with Weka to add machine learning algorithms to the system. The R project adds statistical functionalities as well. Furthermore, KNIME’s range of functions is impressive, with more than 1,000 modules and ready-made application packages. The modules can be further expanded by additional commercial features.

Six properties of modern Business Intelligence

Regardless of the industry in which you operate, you need information systems that evaluate your business data in order to provide you with a basis for decision-making. These systems are commonly referred to as so-called business intelligence (BI). In fact, most BI systems suffer from deficiencies that can be eliminated. In addition, modern BI can partially automate decisions and enable comprehensive analyzes with a high degree of flexibility in use.


Read this article in German:
“Sechs Eigenschaften einer modernen Business Intelligence“


Let us discuss the six characteristics that distinguish modern business intelligence, which mean taking technical tricks into account in detail, but always in the context of a great vision for your own company BI:

1. Uniform database of high quality

Every managing director certainly knows the situation that his managers do not agree on how many costs and revenues actually arise in detail and what the margins per category look like. And if they do, this information is often only available months too late.

Every company has to make hundreds or even thousands of decisions at the operational level every day, which can be made much more well-founded if there is good information and thus increase sales and save costs. However, there are many source systems from the company’s internal IT system landscape as well as other external data sources. The gathering and consolidation of information often takes up entire groups of employees and offers plenty of room for human error.

A system that provides at least the most relevant data for business management at the right time and in good quality in a trusted data zone as a single source of truth (SPOT). SPOT is the core of modern business intelligence.

In addition, other data on BI may also be made available which can be useful for qualified analysts and data scientists. For all decision-makers, the particularly trustworthy zone is the one through which all decision-makers across the company can synchronize.

2. Flexible use by different stakeholders

Even if all employees across the company should be able to access central, trustworthy data, with a clever architecture this does not exclude that each department receives its own views of this data. Many BI systems fail due to company-wide inacceptance because certain departments or technically defined employee groups are largely excluded from BI.

Modern BI systems enable views and the necessary data integration for all stakeholders in the company who rely on information and benefit equally from the SPOT approach.

3. Efficient ways to expand (time to market)

The core users of a BI system are particularly dissatisfied when the expansion or partial redesign of the information system requires too much of patience. Historically grown, incorrectly designed and not particularly adaptable BI systems often employ a whole team of IT staff and tickets with requests for change requests.

Good BI is a service for stakeholders with a short time to market. The correct design, selection of software and the implementation of data flows / models ensures significantly shorter development and implementation times for improvements and new features.

Furthermore, it is not only the technology that is decisive, but also the choice of organizational form, including the design of roles and responsibilities – from the technical system connection to data preparation, pre-analysis and support for the end users.

4. Integrated skills for Data Science and AI

Business intelligence and data science are often viewed and managed separately from each other. Firstly, because data scientists are often unmotivated to work with – from their point of view – boring data models and prepared data. On the other hand, because BI is usually already established as a traditional system in the company, despite the many problems that BI still has today.

Data science, often referred to as advanced analytics, deals with deep immersion in data using exploratory statistics and methods of data mining (unsupervised machine learning) as well as predictive analytics (supervised machine learning). Deep learning is a sub-area of ​​machine learning and is used for data mining or predictive analytics. Machine learning is a sub-area of ​​artificial intelligence (AI).

In the future, BI and data science or AI will continue to grow together, because at the latest after going live, the prediction models flow back into business intelligence. BI will probably develop into ABI (Artificial Business Intelligence). However, many companies are already using data mining and predictive analytics in the company, using uniform or different platforms with or without BI integration.

Modern BI systems also offer data scientists a platform to access high-quality and more granular raw data.

5. Sufficiently high performance

Most readers of these six points will probably have had experience with slow BI before. It takes several minutes to load a daily report to be used in many classic BI systems. If loading a dashboard can be combined with a little coffee break, it may still be acceptable for certain reports from time to time. At the latest, however, with frequent use, long loading times and unreliable reports are no longer acceptable.

One reason for poor performance is the hardware, which can be almost linearly scaled to higher data volumes and more analysis complexity using cloud systems. The use of cloud also enables the modular separation of storage and computing power from data and applications and is therefore generally recommended, but not necessarily the right choice for all companies.

In fact, performance is not only dependent on the hardware, the right choice of software and the right choice of design for data models and data flows also play a crucial role. Because while hardware can be changed or upgraded relatively easily, changing the architecture is associated with much more effort and BI competence. Unsuitable data models or data flows will certainly bring the latest hardware to its knees in its maximum configuration.

6. Cost-effective use and conclusion

Professional cloud systems that can be used for BI systems offer total cost calculators, such as Microsoft Azure, Amazon Web Services and Google Cloud. With these computers – with instruction from an experienced BI expert – not only can costs for the use of hardware be estimated, but ideas for cost optimization can also be calculated. Nevertheless, the cloud is still not the right solution for every company and classic calculations for on-premise solutions are necessary.

Incidentally, cost efficiency can also be increased with a good selection of the right software. Because proprietary solutions are tied to different license models and can only be compared using application scenarios. Apart from that, there are also good open source solutions that can be used largely free of charge and can be used for many applications without compromises.

However, it is wrong to assess the cost of a BI only according to its hardware and software costs. A significant part of cost efficiency is complementary to the aspects for the performance of the BI system, because suboptimal architectures work wastefully and require more expensive hardware than neatly coordinated architectures. The production of the central data supply in adequate quality can save many unnecessary processes of data preparation and many flexible analysis options also make redundant systems unnecessary and lead to indirect savings.

In any case, a BI for companies with many operational processes is always cheaper than no BI. However, if you take a closer look with BI expertise, cost efficiency is often possible.

Interview – Predictive Maintenance and how it can unleash cost savings

Interview with Dr. Kai Goebel, Principal Scientist at PARC, a Xerox Company, about Predictive Maintenance and how it can unleash cost savings.

Dr. Kai Goebel is principal scientist as PARC with more than two decades experience in corporate and government research organizations. He is responsible for leading applied research on state awareness, prognostics and decision-making using data analytics, AI, hybrid methods and physics-base methods. He has also fielded numerous applications for Predictive Maintenance at General Electric, NASA, and PARC for uses as diverse as rocket launchpads, jet engines, and chemical plants.

Data Science Blog: Mr. Goebel, predictive maintenance is not just a hype since industrial companies are already trying to establish this use case of predictive analytics. What benefits do they really expect from it?

Predictive Maintenance is a good example for how value can be realized from analytics. The result of the analytics drives decisions about when to schedule maintenance in advance of an event that might cause unexpected shutdown of the process line. This is in contrast to an uninformed process where the decision is mostly reactive, that is, maintenance is scheduled because equipment has already failed. It is also in contrast to a time-based maintenance schedule. The benefits of Predictive Maintenance are immediately clear: one can avoid unexpected downtime, which can lead to substantial production loss. One can manage inventory better since lead times for equipment replacement can be managed well. One can also manage safety better since equipment health is understood and safety averse situations can potentially be avoided. Finally, maintenance operations will be inherently more efficient as they shift significant time from inspection to mitigation of.

Data Science Blog: What are the most critical success factors for implementing predictive maintenance?

Critical for success is to get the trust of the operator. To that end, it is imperative to understand the limitations of the analytics approach and to not make false performance promises. Often, success factors for implementation hinge on understanding the underlying process and the fault modes reasonably well. It is important to be able to recognize the difference between operational changes and abnormal conditions. It is equally important to recognize rare events reliably while keeping false positives in check.

Data Science Blog: What kind of algorithm does predictive maintenance work with? Do you differentiate between approaches based on classical machine learning and those based on deep learning?

Well, there is no one kind of algorithm that works for Predictive Mantenance everywhere. Instead, one should look at the plurality of all algorithms as tools in a toolbox. Then analyze the problem – how many examples for run-to-failure trajectories are there; what is the desired lead time to report on a problem; what is the acceptable false positive/false negative rate; what are the different fault modes; etc – and use the right kind of tool to do the job. Just because a particular approach (like the one you mentioned in your question) is all the hype right now does not mean it is the right tool for the problem. Sometimes, approaches from what you call “classical machine learning” actually work better. In fact, one should consider approaches even outside the machine learning domain, either as stand-alone approach as in a hybrid configuration. One may also have to invent new methods, for example to perform online learning of the dynamic changes that a system undergoes through its (long) life. In the end, a customer does not care about what approach one is using, only if it solves the problem.

Data Science Blog: There are several providers for predictive analytics software. Is it all about software tools? What makes the difference for having success?

Frequently, industrial partners lament that they have to spend a lot of effort in teaching a new software provider about the underlying industrial processes as well as the equipment and their fault modes. Others are tired of false promises that any kind of data (as long as you have massive amounts of it) can produce any kind of performance. If one does not physically sense a certain modality, no algorithmic magic can take place. In other words, it is not just all about the software. The difference for having success is understanding that there is no cookie cutter approach. And that realization means that one may have to role up the sleeves and to install new instrumentation.

Data Science Blog: What are coming trends? What do you think will be the main topic 2020 and 2021?

Predictive Maintenance is slowly evolving towards Prescriptive Maintenance. Here, one does not only seek to inform about an impending problem, but also what to do about it. Such an approach needs to integrate with the logistics element of an organization to find an optimal decision that trades off several objectives with regards to equipment uptime, process quality, repair shop loading, procurement lead time, maintainer availability, safety constraints, contractual obligations, etc.

Visual Question Answering with Keras – Part 2: Making Computers Intelligent to answer from images

Making Computers Intelligent to answer from images

This is my second blog on Visual Question Answering, in the last blog, I have introduced to VQA, available datasets and some of the real-life applications of VQA. If you have not gone through then I would highly recommend you to go through it. Click here for more details about it.

In this blog post, I will walk through the implementation of VQA in Keras.

You can download the dataset from here: https://visualqa.org/index.html. All my experiments were performed with VQA v2 and I have used a very tiny subset of entire dataset i.e all samples for training and testing from the validation set.

Table of contents:

  1. Preprocessing Data
  2. Process overview for VQA
  3. Data Preprocessing – Images
  4. Data Preprocessing through the spaCy library- Questions
  5. Model Architecture
  6. Defining model parameters
  7. Evaluating the model
  8. Final Thought
  9. References

NOTE: The purpose of this blog is not to get the state-of-art performance on VQA. But the idea is to get familiar with the concept. All my experiments were performed with the validation set only.

Full code on my Github here.


1. Preprocessing Data:

If you have downloaded the dataset then the question and answers (called as annotations) are in JSON format. I have provided the code to extract the questions, annotations and other useful information in my Github repository. All extracted information is stored in .txt file format. After executing code the preprocessing directory will have the following structure.

All text files will be used for training.

 

2. Process overview for VQA:

As we have discussed in previous post visual question answering is broken down into 2 broad-spectrum i.e. vision and text.  I will represent the Neural Network approach to this problem using the Convolutional Neural Network (for image data) and Recurrent Neural Network(for text data). 

If you are not familiar with RNN (more precisely LSTM) then I would highly recommend you to go through Colah’s blog and Andrej Karpathy blog. The concepts discussed in this blogs are extensively used in my post.

The main idea is to get features for images from CNN and features for the text from RNN and finally combine them to generate the answer by passing them through some fully connected layers. The below figure shows the same idea.

 

I have used VGG-16 to extract the features from the image and LSTM layers to extract the features from questions and combining them to get the answer.

3. Data Preprocessing – Images:

Images are nothing but one of the input to our model. But as you already may know that before feeding images to the model we need to convert into the fixed-size vector.

So we need to convert every image into a fixed-size vector then it can be fed to the neural network. For this, we will use the VGG-16 pretrained model. VGG-16 model architecture is trained on millions on the Imagenet dataset to classify the image into one of 1000 classes. Here our task is not to classify the image but to get the bottleneck features from the second last layer.

Hence after removing the softmax layer, we get a 4096-dimensional vector representation (bottleneck features) for each image.

Image Source: https://www.cs.toronto.edu/~frossard/post/vgg16/

 

For the VQA dataset, the images are from the COCO dataset and each image has unique id associated with it. All these images are passed through the VGG-16 architecture and their vector representation is stored in the “.mat” file along with id. So in actual, we need not have to implement VGG-16 architecture instead we just do look up into file with the id of the image at hand and we will get a 4096-dimensional vector representation for the image.

4. Data Preprocessing through the spaCy library- Questions:

spaCy is a free, open-source library for advanced Natural Language Processing (NLP) in Python. As we have converted images into a fixed 4096-dimensional vector we also need to convert questions into a fixed-size vector representation. For installing spaCy click here

You might know that for training word embeddings in Keras we have a layer called an Embedding layer which takes a word and embeds it into a higher dimensional vector representation. But by using the spaCy library we do not have to train the get the vector representation in higher dimensions.

 

This model is actually trained on billions of tokens of the large corpus. So we just need to call the vector method of spaCy class and will get vector representation for word.

After fitting, the vector method on tokens of each question will get the 300-dimensional fixed representation for each word.

5. Model Architecture:

In our problem the input consists of two parts i.e an image vector, and a question, we cannot use the Sequential API of the Keras library. For this reason, we use the Functional API which allows us to create multiple models and finally merge models.

The below picture shows the high-level architecture idea of submodules of neural network.

After concatenating the 2 different models the summary will look like the following.

The below plot helps us to visualize neural network architecture and to understand the two types of input:

 

6. Defining model parameters:

The hyperparameters that we are going to use for our model is defined as follows:

If you know what this parameter means then you can play around it and can get better results.

Time Taken: I used the GPU on https://colab.research.google.com and hence it took me approximately 2 hours to train the model for 5 epochs. However, if you train it on a PC without GPU, it could take more time depending on the configuration of your machine.

7. Evaluating the model:

Since I have used the very small dataset for performing these experiments I am not able to get very good accuracy. The below code will calculate the accuracy of the model.

 

Since I have trained a model multiple times with different parameters you will not get the same accuracy as me. If you want you can directly download mode.h5 file from my google drive.

 

8. Final Thoughts:

One of the interesting thing about VQA is that it a completely new field. So there is absolutely no end to what you can do to solve this problem. Below are some tips while replicating the code.

  1. Start with a very small subset of data: When you start implementing I suggest you start with a very small amount of data. Because once you are ready with the whole setup then you can scale it any time.
  2. Understand the code: Understanding code line by line is very much helpful to match your theoretical knowledge. So for that, I suggest you can take very few samples(maybe 20 or less) and run a small chunk (2 to 3 lines) of code to get the functionality of each part.
  3. Be patient: One of the mistakes that I did while starting with this project was to do everything at one go. If you get some error while replicating code spend 4 to 5 days harder on that. Even after that if you won’t able to solve, I would suggest you resume after a break of 1 or 2 days. 

VQA is the intersection of NLP and CV and hopefully, this project will give you a better understanding (more precisely practically) with most of the deep learning concepts.

If you want to improve the performance of the model below are few tips you can try:

  1. Use larger datasets
  2. Try Building more complex models like Attention, etc
  3. Try using other pre-trained word embeddings like Glove 
  4. Try using a different architecture 
  5. Do more hyperparameter tuning

The list is endless and it goes on.

In the blog, I have not provided the complete code you can get it from my Github repository.

9. References:

  1. https://blog.floydhub.com/asking-questions-to-images-with-deep-learning/
  2. https://tryolabs.com/blog/2018/03/01/introduction-to-visual-question-answering/
  3. https://github.com/sominwadhwa/vqamd_floyd

Visual Question Answering with Keras – Part 1

This is Part I of II of the Article Series Visual Question Answering with Keras

Making Computers Intelligent to answer from images

If we look closer in the history of Artificial Intelligence (AI), the Deep Learning has gained more popularity in the recent years and has achieved the human-level performance in the tasks such as Speech Recognition, Image Classification, Object Detection, Machine Translation and so on. However, as humans, not only we but also a five-year child can normally perform these tasks without much inconvenience. But the development of such systems with these capabilities has always considered an ambitious goal for the researchers as well as for developers.

In this series of blog posts, I will cover an introduction to something called VQA (Visual Question Answering), its available datasets, the Neural Network approach for VQA and its implementation in Keras and the applications of this challenging problem in real life. 

Table of Contents:

1 Introduction

2 What is exactly Visual Question Answering?

3 Prerequisites

4 Datasets available for VQA

4.1 DAQUAR Dataset

4.2 CLEVR Dataset

4.3 FigureQA Dataset

4.4 VQA Dataset

5 Real-life applications of VQA

6 Conclusion

 

  1. Introduction:

Let’s say you are given a below picture along with one question. Can you answer it?

I expect confidently you all say it is the Kitchen without much inconvenience which is also the right answer. Even a five-year child who just started to learn things might answer this question correctly.

Alright, but can you write a computer program for such type of task that takes image and question about the image as an input and gives us answer as output?

Before the development of the Deep Neural Network, this problem was considered as one of the difficult, inconceivable and challenging problem for the AI researcher’s community. However, due to the recent advancement of Deep Learning the systems are capable of answering these questions with the promising result if we have a required dataset.

Now I hope you have got at least some intuition of a problem that we are going to discuss in this series of blog posts. Let’s try to formalize the problem in the below section.

  1. What is exactly Visual Question Answering?:

We can define, “Visual Question Answering(VQA) is a system that takes an image and natural language question about the image as an input and generates natural language answer as an output.”

VQA is a research area that requires an understanding of vision(Computer Vision)  as well as text(NLP). The main beauty of VQA is that the reasoning part is performed in the context of the image. So if we have an image with the corresponding question then the system must able to understand the image well in order to generate an appropriate answer. For example, if the question is the number of persons then the system must able to detect faces of the persons. To answer the color of the horse the system need to detect the objects in the image. Many of these common problems such as face detection, object detection, binary object classification(yes or no), etc. have been solved in the field of Computer Vision with good results.

To summarize a good VQA system must be able to address the typical problems of CV as well as NLP.

To get a better feel of VQA you can try online VQA demo by CloudCV. You just go to this link and try uploading the picture you want and ask the related question to the picture, the system will generate the answer to it.

 

  1. Prerequisites:

In the next post, I will walk you through the code for this problem using Keras. So I assume that you are familiar with:

  1. Fundamental concepts of Machine Learning
  2. Multi-Layered Perceptron
  3. Convolutional Neural Network
  4. Recurrent Neural Network (especially LSTM)
  5. Gradient Descent and Backpropagation
  6. Transfer Learning
  7. Hyperparameter Optimization
  8. Python and Keras syntax
  1. Datasets available for VQA:

As you know problems related to the CV or NLP the availability of the dataset is the key to solve the problem. The complex problems like VQA, the dataset must cover all possibilities of questions answers in real-world scenarios. In this section, I will cover some of the datasets available for VQA.

4.1 DAQUAR Dataset:

The DAQUAR dataset is the first dataset for VQA that contains only indoor scenes. It shows the accuracy of 50.2% on the human baseline. It contains images from the NYU_Depth dataset.

Example of DAQUAR dataset

Example of DAQUAR dataset

The main disadvantage of DAQUAR is the size of the dataset is very small to capture all possible indoor scenes.

4.2 CLEVR Dataset:

The CLEVR Dataset from Stanford contains the questions about the object of a different type, colors, shapes, sizes, and material.

It has

  • A training set of 70,000 images and 699,989 questions
  • A validation set of 15,000 images and 149,991 questions
  • A test set of 15,000 images and 14,988 questions

Image Source: https://cs.stanford.edu/people/jcjohns/clevr/?source=post_page

 

4.3 FigureQA Dataset:

FigureQA Dataset contains questions about the bar graphs, line plots, and pie charts. It has 1,327,368 questions for 100,000 images in the training set.

4.4 VQA Dataset:

As comapred to all datasets that we have seen so far VQA dataset is relatively larger. The VQA dataset contains open ended as well as multiple choice questions. VQA v2 dataset contains:

  • 82,783 training images from COCO (common objects in context) dataset
  • 40, 504 validation images and 81,434 validation images
  • 443,757 question-answer pairs for training images
  • 214,354 question-answer pairs for validation images.

As you might expect this dataset is very huge and contains 12.6 GB of training images only. I have used this dataset in the next post but a very small subset of it.

This dataset also contains abstract cartoon images. Each image has 3 questions and each question has 10 multiple choice answers.

  1. Real-life applications of VQA:

There are many applications of VQA. One of the famous applications is to help visually impaired people and blind peoples. In 2016, Microsoft has released the “Seeing AI” app for visually impaired people to describe the surrounding environment around them. You can watch this video for the prototype of the Seeing AI app.

Another application could be on social media or e-commerce sites. VQA can be also used for educational purposes.

  1. Conclusion:

I hope this explanation will give you a good idea of Visual Question Answering. In the next blog post, I will walk you through the code in Keras.

If you like my explanations, do provide some feedback, comments, etc. and stay tuned for the next post.

Attribution Models in Marketing

Attribution Models

A Business and Statistical Case

INTRODUCTION

A desire to understand the causal effect of campaigns on KPIs

Advertising and marketing costs represent a huge and ever more growing part of the budget of companies. Studies have found out this share is as high as 10% and increases with the size of companies (CMO study by American Marketing Association and Duke University, 2017). Measuring precisely the impact of a specific marketing campaign on the sales of a company is a critical step towards an efficient allocation of this budget. Would the return be higher for an euro spent on a Facebook ad, or should we better spend it on a TV spot? How much should I spend on Twitter ads given the volume of sales this channel is responsible for?

Attribution Models have lately received great attention in Marketing departments to answer these issues. The transition from offline to online marketing methods has indeed permitted the collection of multiple individual data throughout the whole customer journey, and  allowed for the development of user-centric attribution models. In short, Attribution Models use the information provided by Tracking technologies such as Google Analytics or Webtrekk to understand customer journeys from the first click on a Facebook ad to the final purchase and adequately ponderate the different marketing campaigns encountered depending on their responsibility in the final conversion.

Issues on Causal Effects

A key question then becomes: how to declare a channel is responsible for a purchase? In other words, how can we isolate the causal effect or incremental value of a campaign ?

          1. A/B-Tests

One method to estimate the pure impact of a campaign is the design of randomized experiments, wherein a control and treated groups are compared.  A/B tests belong to this broad category of randomized methods. Provided the groups are a priori similar in every aspect except for the treatment received, all subsequent differences may be attributed solely to the treatment. This method is typically used in medical studies to assess the effect of a drug to cure a disease.

Main practical issues regarding Randomized Methods are:

  • Assuring that control and treated groups are really similar before treatment. Uually a random assignment (i.e assuring that on a relevant set of observable variables groups are similar) is realized;
  • Potential spillover-effects, i.e the possibility that the treatment has an impact on the non-treated group as well (Stable unit treatment Value Assumption, or SUTVA in Rubin’s framework);
  • The costs of conducting such an experiment, and especially the costs linked to the deliberate assignment of individuals to a group with potentially lower results;
  • The number of such experiments to design if multiple treatments have to be measured;
  • Difficulties taking into account the interaction effects between campaigns or the effect of spending levels. Indeed, usually A/B tests are led by cutting off temporarily one campaign entirely and measuring the subsequent impact on KPI’s compared to the situation where this campaign is maintained;
  • The dynamical reproduction of experiments if we assume that treatment effects may change over time.

In the marketing context, multiple campaigns must be tested in a dynamical way, and treatment effect is likely to be heterogeneous among customers, leading to practical issues in the lauching of A/B tests to approximate the incremental value of all campaigns. However, sites with a lot of traffic and conversions can highly benefit from A/B testing as it provides a scientific and straightforward way to approximate a causal impact. Leading companies such as Uber, Netflix or Airbnb rely on internal tools for A/B testing automation, which allow them to basically test any decision they are about to make.

References:

Books:

Experiment!: Website conversion rate optimization with A/B and multivariate testing, Colin McFarland, ©2013 | New Riders  

A/B testing: the most powerful way to turn clicks into customers. Dan Siroker, Pete Koomen; Wiley, 2013.

Blogs:

https://eng.uber.com/xp

https://medium.com/airbnb-engineering/growing-our-host-community-with-online-marketing-9b2302299324

Study:

https://cmosurvey.org/wp-content/uploads/sites/15/2018/08/The_CMO_Survey-Results_by_Firm_and_Industry_Characteristics-Aug-2018.pdf

        2. Attribution models

Attribution Models do not demand to create an experimental setting. They take into account existing data and derive insights from the variability of customer journeys. One key difficulty is then to differentiate correlation and causality in the links observed between the exposition to campaigns and purchases. Indeed, selection effects may bias results as exposure to campaigns is usually dependant on user-characteristics and thus may not be necessarily independant from the customer’s baseline conversion probabilities. For example, customers purchasing from a discount price comparison website may be intrinsically different from customers buying from FB ad and this a priori difference may alone explain post-exposure differences in purchasing bahaviours. This intrinsic weakness must be remembered when interpreting Attribution Models results.

                          2.1 General Issues

The main issues regarding the implementation of Attribution Models are linked to

  • Causality and fallacious reasonning, as most models do not take into account the aforementionned selection biases.
  • Their difficult evaluation. Indeed, in almost all attribution models (except for those based on classification, where the accuracy of the model can be computed), the additionnal value brought by the use of a given attribution models cannot be evaluated using existing historical data. This additionnal value can only be approximated by analysing how the implementation of the conclusions of the attribution model have impacted a given KPI.
  • Tracking issues, leading to an uncorrect reconstruction of customer journeys
    • Cross-device journeys: cross-device issue arises from the use of different devices throughout the customer journeys, making it difficult to link datapoints. For example, if a customer searches for a product on his computer but later orders it on his mobile, the AM would then mistakenly consider it an order without prior campaign exposure. Though difficult to measure perfectly, the proportion of cross-device orders can approximate 20-30%.
    • Cookies destruction makes it difficult to track the customer his the whole journey. Both regulations and consumers’ rising concerns about data privacy issues mitigate the reliability and use of cookies.1 – From 2002 on, the EU has enacted directives concerning privacy regulation and the extended use of cookies for commercial targeting purposes, which have highly impacted marketing strategies, such as the ‘Privacy and Electronic Communications Directive’ (2002/58/EC). A research was conducted and found out that the adoption of this ‘Privacy Directive’ had led to 64% decrease in advertising methods compared to the rest of the world (Goldfarb et Tucker (2011)). The effect was stronger for generalized sites (Yahoo) than for specialized sites.2 – Users have grown more and more conscious of data privacy issues and have adopted protective measures concerning data privacy, such as automatic destruction of cookies after a session is ended, or simply giving away less personnal information (Goldfarb et Tucker (2012) ) .Valuable user information may be lost, though tracking technologies evolution have permitted to maintain tracking by other means. This issue may be particularly important in countries highly concerned with data privacy issues such as Germany.
    • Offline/Online bridge: an Attribution Model should take into account all campaigns to draw valuable insights. However, the exposure to offline campaigns (TV, newspapers) are difficult to track at the user level. One idea to tackle this issue would be to estimate the proportion of conversions led by offline campaigns through AB testing and deduce this proportion from the credit assigned to the online campaigns accounted for in the Attribution Model.
    • Touch point information available: clicks are easy to follow but irrelevant to take into account the influence of purely visual campaigns such as display ads or video.

                          2.2 Today’s main practices

Two main families of Attribution Models exist:

  • Rule-Based Attribution Models, which have been used for in the last decade but from which companies are gradualy switching.

Attribution depends on the individual journeys that have led to a purchase and is solely based on the rank of the campaign in the journey. Some models focus on a single touch points (First Click, Last Click) while others account for multi-touch journeys (Bathtube, Linear). It can be calculated at the customer level and thus doesn’t require large amounts of data points. We can distinguish two sub-groups of rule-based Attribution Models:

  • One Touch Attribution Models attribute all credit to a single touch point. The First-Click model attributes all credit for a converion to the first touch point of the customer journey; last touch attributes all credit to the last campaign.
  • Multi-touch Rule-Based Attribution Models incorporate information on the whole customer journey are thus an improvement compared to one touch models. To this family belong Linear model where credit is split equally between all channels, Bathtube model where 40% of credit is given to first and last clicks and the remaining 20% is distributed equally between the middle channels, or time-decay models where credit assigned to a click diminishes as the time between the click and the order increases..

The main advantages of rule-based models is their simplicity and cost effectiveness. The main problems are:

– They are a priori known and can thus lead to optimization strategies from competitors
– They do not take into account aggregate intelligence on customer journeys and actual incremental values.
– They tend to bias (depending on the model chosen) channels that are over-represented at the beggining or end of the funnel, according to theoretical assumptions that have no observationnal back-ups.

  • Data-Driven Attribution Models

These models take into account the weaknesses of rule-based models and make a relevant use of available data. Being data-driven, following attribution models cannot be computed using single user level data. On the contrary values are calculated through data aggregation and thus require a certain volume of customer journey information.

References:

https://dspace.mit.edu/handle/1721.1/64920

 

        3. Data-Driven Attribution Models in practice

                          3.1 Issues

Several issues arise in the computation of campaigns individual impact on a given KPI within a data-driven model.

  • Selection biases: Exposure to certain types of advertisement is usually highly correlated to non-observable variables which are in turn correlated to consumption practices. Differences in the behaviour of users exposed to different campaigns may thus only be driven by core differences in conversion probabilities between groups whether than by the campaign effect.
  • Complementarity: it may be that campaigns A and B only have an effect when combined, so that measuring their individual impact would lead to misleading conclusions. The model could then try to assess the effect of combinations of campaigns on top of the effect of individual campaigns. As the number of possible non-ordered combinations of k campaigns is 2k, it becomes clear that inclusing all possible combinations would however be time-consuming.
  • Order-sensitivity: The effect of a campaign A may depend on the place where it appears in the customer journey, meaning the rank of a campaign and not merely its presence could be accounted for in the model.
  • Relative Order-sensitivity: it may be that campaigns A and B only have an effect when one is exposed to campaign A before campaign B. If so, it could be useful to assess the effect of given combinations of campaigns as well. And this for all campaigns, leading to tremendous numbers of possible combinations.
  • All previous phenomenon may be present, increasing even more the potential complexity of a comprehensive Attribution Model. The number of all possible ordered combination of k campaigns is indeed :

 

                          3.2 Main models

                                  A) Logistic Regression and Classification models

If non converting journeys are available, Attribition Model can be shaped as a simple classification issue. Campaign types or campaigns combination and volume of campaign types can be included in the model along with customer or time variables. As we are interested in inference (on campaigns effect) whether than prediction, a parametric model should be used, such as Logistic Regression. Non paramatric models such as Random Forests or Neural Networks can also be used though the interpretation of campaigns value would be more difficult to derive from the model results.

A common pitfall is the usual issue of spurious correlations on one hand and the correct interpretation of coefficients in business terms.

An advantage if the possibility to evaluate the relevance of the model using common model validation methods to evaluate its predictive power (validation set \ AUC \pseudo R squared).

                                  B) Shapley Value

Theory

The Shapley Value is based on a Game Theory framework and is named after its creator, the Nobel Price Laureate Lloyd Shapley. Initially meant to calculate the marginal contribution of players in cooperative games, the model has received much attention in research and industry and has lately been applied to marketing issues. This model is typically used by Google Adords and other ad bidding vendors. Campaigns or marketing channels are in this model seen as compementary players looking forward to increasing a given KPI.
Contrarily to Logistic Regressions, it is a non-parametric model. Contrarily to Markov Chains, all results are built using existing journeys, and not simulated ones.

Channels are considered to enter the game sequentially under a certain joining order. Shapley value try to The Shapley value of channel i is the weighted sum of the marginal values that channel i adds to all possible coalitions that don’t contain channel i.
In other words, the main logic is to analyse the difference of gains when a channel i is added after a coalition Ck of k channels, k<=n. We then sum all the marginal contributions over all possible ordered combination Ck of all campaigns excluding i, with k<=n-1.

Subsets framework

A first an most usual way to compute the Shapley Vaue is to consider that when a channel enters coalition, its additionnal value is the same irrelevant of the order in which previous channels have appeared. In other words, journeys (A>B>C) and (B>A>C) trigger the same gains.
Shapley value is computed as the gains associated to adding a channel i to a subset of channels, weighted by the number of (ordered) sequences that the (unordered) subset represents, summed up on all possible subsets of the total set of campaigns where the channel i is not present.
The Shapley value of the channel ???????? is then:

where |S| is the number of campaigns of a coalition S and the sum extends over all subsets S that do not not contain channel j. ????(????)  is the value of the coalition S and ????(???? ∪ {????????})  the value of the coalition formed by adding ???????? to coalition S. ????(???? ∪ {????????}) − ????(????) is thus the marginal contribution of channel ???????? to the coalition S.

The formula can be rewritten and understood as:

This method is convenient when data on the gains of on all possible permutations of all unordered k subsets of the n campaigns are available. It is also more convenient if the order of campaigns prior to the introduction of a campaign is thought to have no impact.

Ordered sequences

Let us define ????((A>B)) as the value of the sequence A then B. What is we let ????((A>B)) be different from ????((B>A)) ?
This time we would need to sum over all possible permutation of the S campaigns present before  ???????? and the N-(S+1) campaigns after ????????. Doing so we will sum over all possible orderings (i.e all permutations of the n campaigns of the grand coalition containing all campaigns) and we can remove the permutation coefficient s!(p-s+1)!.

This method is convenient when the order of channels prior to and after the introduction of another channel is assumed to have an impact. It is also necessary to possess data for all possible permutations of all k subsets of the n campaigns, and not only on all (unordered) k-subsets of the n campaigns, k<=n. In other words, one must know the gains of A, B, C, A>B, B>A, etc. to compute the Shapley Value.

Differences between the two approaches

We simulate an ordered case where the value for each ordered sequence k for k<=3 is known. We compare it to the usual Shapley value calculated based on known gains of unordered subsets of campaigns. So as to compare relevant values, we have built the gains matrix so that the gains of a subset A, B i.e  ????({B,A}) is the average of the gains of ordered sequences made up with A and B (assuming the number of journeys where A>B equals the number of journeys where B>A, we have ????({B,A})=0.5( ????((A>B)) + ????((B>A)) ). We let the value of the grand coalition be different depending on the order of campaigns-keeping the constraints that it averages to the value used for the unordered case.

Note: mvA refers to the marginal value of A in a given sequence.
With traditionnal unordered coalitions:

With ordered sequences used to compute the marginal values:

 

We can see that the two approaches yield very different results. In the unordered case, the Shapley Value campaign C is the highest, culminating at 20, while A and B have the same Shapley Value mvA=mvB=15. In the ordered case, campaign A has the highest Shapley Value and all campaigns have different Shapley Values.

This example illustrates the inherent differences between the set and sequences approach to Shapley values. Real life data is more likely to resemble the ordered case as conversion probabilities may for any given set of campaigns be influenced by the order through which the campaigns appear.

Advantages

Shapley value has become popular in allocation problems in cooperative games because it is the unique allocation which satisfies different axioms:

  • Efficiency: Shaple Values of all channels add up to the total gains (here, orders) observed.
  • Symmetry: if channels A and B bring the same contribution to any coalition of campaigns, then their Shapley Value i sthe same
  • Null player: if a channel brings no additionnal gains to all coalitions, then its Shapley Value is zero
  • Strong monotony: the Shapley Value of a player increases weakly if all its marginal contributions increase weakly

These properties make the Shapley Value close to what we intuitively define as a fair attribution.

Issues

  • The Shapley Value is based on combinatory mathematics, and the number of possible coalitions and ordered sequences becomes huge when the number of campaigns increases.
  • If unordered, the Shapley Value assumes the contribution of campaign A is the same if followed by campaign B or by C.
  • If ordered, the number of combinations for which data must be available and sufficient is huge.
  • Channels rarely present or present in long journeys will be played down.
  • Generally, gains are supposed to grow with the number of players in the game. However, it is plausible that in the marketing context a journey with a high number of channels will not necessarily bring more orders than a journey with less channels involved.

References:

R package: GameTheoryAllocation

Article:
Zhao & al, 2018 “Shapley Value Methods for Attribution Modeling in Online Advertising “
https://link.springer.com/content/pdf/10.1007/s13278-017-0480-z.pdf
Courses: https://www.lamsade.dauphine.fr/~airiau/Teaching/CoopGames/2011/coopgames-7%5b8up%5d.pdf
Blogs: https://towardsdatascience.com/one-feature-attribution-method-to-supposedly-rule-them-all-shapley-values-f3e04534983d

                                  B) Markov Chains

Markov Chains are used to model random processes, i.e events that occur in a sequential manner and in such a way that the probability to move to a certain state only depends on the past steps. The number of previous steps that are taken into account to model the transition probability is called the memory parameter of the sequence, and for the model to have a solution must be comprised between 0 and 4. A Markov Chain process is thus defined entirely by its Transition Matrix and its initial vector (i.e the starting point of the process).

Markov Chains are applied in many scientific fields. Typically, they are used in weather forecasting, with the sequence of Sunny and Rainy days following a Markov Process of memory parameter 0, so that for each given day the probability that the next day will be rainy or sunny only depends on the weather of the current day. Other applications can be found in sociology to understand the dynamics of social classes intergenerational reproduction. To get more both mathematical and applied illustration, I recommend the reading of this course.

In the marketing context, Markov Chains are an interesting way to model the conversion funnel. To go from the from the Markov Model to the Attribution logic, we calculate the Removal Effect of each channel, i.e the difference in conversions that happen if the channel is removed. Please read below for an introduction to the methodology.

The first step in a Markov Chains Attribution Model is to build the transition matrix that captures the transition probabilities between the campaigns accross existing customer journeys. This Matrix is to be read as a “From state A to state B” table, from the left to the right. A first difficulty is finding the right memory parameter to use. A large memory parameter would allow to take more into account interraction effects within the conversion funnel but would lead to increased computationnal time, a non-readable transition matrix, and be more sensitive to noisy data. Please note that this transition matrix provides useful information on the conversion funnel and on the relationships between campaigns and can be used as such as an analytical tool. I suggest the clear and easily R code which can be found here or here.

Here is an illustration of a Markov Chain with memory Parameter of 0: the probability to go to a certain campaign B in the next step only depend on the campaign we are currently at:

The associated Transition Matrix is then (with null probabilities left as Blank):

The second step is  to compute the actual responsibility of a channel in total conversions. As mentionned above, the main philosophy to do so is to calculate the Removal Effect of each channel, i.e the changes in the number of conversions when a channel is entirely removed. All customer journeys which went through this channel are settled out to be unsuccessful. This calculation is done by applying the transition matrix with and without the removed channels to an initial vector that contains the number of desired simulations.

Building on our current example, we can then settle an initial vector with the desired number of simulations, e.g 10 000:

 

It is possible at this stage to add a constraint on the maximum number of times the matrix is applied to the data, i.e on the maximal number of campaigns a simulated journey is allowed to have.

Advantages

  • The dynamic journey is taken into account, as well as the transition between two states. The funnel is not assumed to be linear.
  • It is possile to build a conversion graph that maps the customer journey provides valuable insights.
  • It is possible to evaluate partly the accuracy of the Attribution Model based on Markov Chains. It is for example possible to see how well the transition matrix help predict the future by analysing the number of correct predictions at any given step over all sequences.

Disadvantages

  • It can be somewhat difficult to set the memory parameter. Complementarity effects between channels are not well taken into account if the memory is low, but a parameter too high will lead to over-sensitivity to noise in the data and be difficult to implement if customer journeys tend to have a number of campaigns below this memory parameter.
  • Long journeys with different channels involved will be overweighted, as they will count many times in the Removal Effect.  For example, if there are n-1 channels in the customer journey, this journey will be considered as failure for the n-1 channel-RE. If the volume effects (i.e the impact of the overall number of channels in a journey, irrelevant from their type° are important then results may be biased.

References:

R package: ChannelAttribution

Git:

https://github.com/MatCyt/Markov-Chain/blob/master/README.md

Course:

https://www.ssc.wisc.edu/~jmontgom/markovchains.pdf

Article:

“Mapping the Customer Journey: A Graph-Based Framework for Online Attribution Modeling”; Anderl, Eva and Becker, Ingo and Wangenheim, Florian V. and Schumann, Jan Hendrik, 2014. Available at SSRN: https://ssrn.com/abstract=2343077 or http://dx.doi.org/10.2139/ssrn.2343077

“Media Exposure through the Funnel: A Model of Multi-Stage Attribution”, Abhishek & al, 2012

“Multichannel Marketing Attribution Using Markov Chains”, Kakalejčík, L., Bucko, J., Resende, P.A.A. and Ferencova, M. Journal of Applied Management and Investments, Vol. 7 No. 1, pp. 49-60.  2018

Blogs:

https://analyzecore.com/2016/08/03/attribution-model-r-part-1

https://analyzecore.com/2016/08/03/attribution-model-r-part-2

                          3.3 To go further: Tackling selection biases with Quasi-Experiments

Exposure to certain types of advertisement is usually highly correlated to non-observable variables. Differences in the behaviour of users exposed to different campaigns may thus only be driven by core differences in converison probabilities between groups whether than by the campaign effect. These potential selection effects may bias the results obtained using historical data.

Quasi-Experiments can help correct this selection effect while still using available observationnal data.  These methods recreate the settings on a randomized setting. The goal is to come as close as possible to the ideal of comparing two populations that are identical in all respects except for the advertising exposure. However, populations might still differ with respect to some unobserved characteristics.

Common quasi-experimental methods used for instance in Public Policy Evaluation are:

  • Discontinuity Regressions
  • Matching Methods, such as Exact Matching,  Propensity-score matching or k-nearest neighbourghs.

References:

Article:

“Towards a digital Attribution Model: Measuring the impact of display advertising on online consumer behaviour”, Anindya Ghose & al, MIS Quarterly Vol. 40 No. 4, pp. 1-XX, 2016

https://pdfs.semanticscholar.org/4fa6/1c53f281fa63a9f0617fbd794d54911a2f84.pdf

        4. First Steps towards a Practical Implementation

Identify key points of interests

  • Identify the nature of touchpoints available: is the data based on clicks? If so, is there a way to complement the data with A/B tests to measure the influence of ads without clicks (display, video) ? For example, what happens to sales when display campaign is removed? Analysing this multiplier effect would give the overall responsibility of display on sales, to be deduced from current attribution values given to click-based channels. More interestingly, what is the impact of the removal of display campaign on the occurences of click-based campaigns ? This would give us an idea of the impact of display ads on the exposure to each other campaigns, which would help correct the attribution values more precisely at the campaign level.
  • Define the KPI to track. From a pure Marketing perspective, looking at purchases may be sufficient, but from a financial perspective looking at profits, though a bit more difficult to compute, may drive more interesting results.
  • Define a customer journey. It may seem obvious, but the notion needs to be clarified at first. Would it be defined by a time limit? If so, which one? Does it end when a conversion is observed? For example, if a customer makes 2 purchases, would the campaigns he’s been exposed to before the first order still be accounted for in the second order? If so, with a time decay?
  • Define the research framework: are we interested only in customer journeys which have led to conversions or in all journeys? Keep in mind that successful customer journeys are a non-representative sample of customer journeys. Models built on the analysis of biased samples may be conservative. Take an extreme example: 80% of customers who see campaign A buy the product, VS 1% for campaign B. However, campaign B exposure is great and 100 Million people see it VS only 1M for campaign A. An Attribution Model based on successful journeys will give higher credit to campaign B which is an auguable conclusion. Taking into account costs per campaign (in the case where costs are calculated by clicks) may of course tackle this issue partly, as campaign A could then exhibit higher returns, but a serious fallacious reasonning is at stake here.

Analyse the typical customer journey    

  • Performing a duration analysis on the data may help you improve the definition of the customer journey to be used by your organization. After which days are converison probabilities null? Should we consider the effect of campaigns disappears after x days without orders? For example, if 99% of orders are placed in the 30 days following a first click, it might be interesting to define the customer journey as a 30 days time frame following the first oder.
  • Look at the distribution of the number of campaigns in a typical journey. If you choose to calculate the effect of campaigns interraction in your Attribution Model, it may indeed help you determine the maximum number of campaigns to be included in a combination. Indeed, you may not need to assess the impact of channel combinations with above than 4 different channels if 95% of orders are placed after less then 4 campaigns.
  • Transition matrixes: what if a campaign A systematically leads to a campaign B? What happens if we remove A or B? These insights would give clues to ask precise questions for a latter AB test, for example to find out if there is complementarity between channels A and B – (implying none should be removed) or mere substitution (implying one can be given up).
  • If conversion rates are available: it can be interesting to perform a survival analysis i.e to analyse the likelihood of conversion based on duration since first click. This could help us excluse potential outliers or individuals who have very low conversion probabilities.

Summary

Attribution is a complex topic which will probably never be definitively solved. Indeed, a main issue is the difficulty, or even impossibility, to evaluate precisely the accuracy of the attribution model that we’ve built. Attribution Models should be seen as a good yet always improvable approximation of the incremental values of campaigns, and be presented with their intrinsinc limits and biases.

Introduction to ROC Curve

The abbreviation ROC stands for Receiver Operating Characteristic. Its main purpose is to illustrate the diagnostic ability of classifier as the discrimination threshold is varied. It was developed during World War II when Radar operators had to decide if the blip on the screen is an enemy target, a friendly ship or just a noise.  For these purposes they measured the ability of a radar receiver operator to make these important distinctions, which was called the Receiver Operating Characteristic.

Later it was found useful in interpreting medical test results and then in Machine learning classification problems. In order to get an introduction to binary classification and terms like ‘precision’ and ‘recall’ one can look into my earlier blog  here.

True positive rate and false positive rate

Let’s imagine a situation where a fire alarm is installed in a kitchen. The alarm is supposed to emit a sound in case fire smoke is detected in the room. Unfortunately, there is a lot of cooking done in the kitchen and the alarm may trigger the sound too often. Thus, instead of serving a purpose the alarm becomes a nuisance due to a large number of false alarms. In statistical terms these types of errors are called type 1 errors, or false positives.

One way to deal with this problem is to simply decrease sensitivity of the device. We do this by increasing the trigger threshold at the alarm setting. But then, not every alarm should have the same threshold setting. Consider the same type of device but kept in a bedroom. With high threshold, the device might miss smoke from a real short-circuit in the wires which poses a real danger of fire. This kind of failure is called Type 2 error or a false negative. Although the two devices are the same, different types of threshold settings are optimal for different circumstances.

To specify this more formally, let us describe the performance of a binary classifier at a particular threshold by the following parameters:

 

These parameters take different values at different thresholds. Hence, they define the performance of the classifier at particular threshold. But we want to examine in overall how good a classifier is. Fortunately, there is a way to do that. We plot the True Positive Rate (TPR) and False Positive rate (FPR) at different thresholds and this plot is called ROC curve.

Let’s try to understand this with an example.

A case with a distinct population distribution

Let’s suppose there is a disease which can be identified with deficiency of some parameter (maybe a certain vitamin). The distribution of population with this disease has a mean vitamin concentration sharply distinct from the mean of a healthy population, as shown below.

This is result of dummy data simulating population of 2000 people,the link to the code is given  in the end of this blog.  As the two populations are distinctly separated (there is no  overlap between the two distributions), we can expect that a classifier would have an easy job distinquishing healthy from sick people. We can run a logistic regression classifier with a threshold of .5 and be 100% succesful in detecting the decease.

The confusion matrix may look something like this.

In this ideal case with a threshold  of  .5 we do not make a single wrong classification. The True positive rate and False positive rate are 1 and 0, respectively. But we can shift the threshold. In that case, we will  get different confusion matrices. First we plot threshold vs. TPR.

We see for most values of threshold the TPR is close to 1 which again proves data is easy to classify and the classifier is returning high probabilities  for the most of positives .

Similarly Let’s plot threshold vs. FPR.

For most of the data points FPR is close to zero. This is also good. Now its time to plot the ROC curve using these results (TPR vs FPR).

Let’s try to interpret  the results,  all the points lie on line x=0 and y=1, it means for all the points FPR is zero or TPR is one, making  the curve a square. which means the classifier does perfectly well.

Case with overlapping  population distribution

The above example was about a perfect classifer. However, life is often not so easy. Now let us consider another more realistic situation in which the parameter distribution of the population is not as distinct as in the previous case. Rather, the mean of the parameter with healthy and not healthy datapoints are close and the distributions overlap, as shown in the next figure.

If we set the threshold to 0.5, the confusion matrix may look like this.

Now, any new choice of threshold location will affect both false positives and false negatives. In fact, there is a trade-off. If we shift the threshold with the goal to reduce false negatives, false positives will increase. If we move the threshold to the other direction and reduce false positive, false negatives will increase.

The plots (TPR vs Threshold) , (FPR vs Threshold) are shown below

If we plot the ROC curve from these results, it looks like this:

From the curve we see the classifier does not perform as well as the earlier one.

What else can be infered from this curve? We first need to understand what the diagonal in this plot represent. The diagonal represents ‘Line of no discrimination’, which we obtain if we randomly guess. This is the ROC curve for the worst possible classifier. Therefore, by comparing the obtained ROC curve with the diagonal, we see how much better our classifer is from random guessing.

The further away ROC curve from the diagonal is (the closest it is to the top left corner) , better the classifier is.

Area Under the curve

The overall performance of the classifier is given by the area under the ROC curve and is usually denoted as AUC. Since TPR and FPR lie within the range of 0 to 1, the AUC also assumes values between 0 and 1. The higher the value of AUC, the better is the overall performance of the classifier.

Let’s see this for the two different distributions which we saw earlier.

As we know the classifier had worked perfectly in the first case with points at (0,1) the area under the curve is 1 which is perfect. In the latter case the classifier was not able to perform as good, the ROC curve is between the diagonal and left hand corner. The AUC as we can see is less than 1.

Some other general characteristics

There are still few points that needs to be discussed on a General ROC curve

  • The ROC curve does not provide information about the actual values of thresholds used for the classifier.
  • Performance of different classifiers can be compared using the AUC of different Classifier. The larger the AUC, the better the classifier.
  • The vertical distance of the ROC curve from the no discrimination line gives a measure of ‘INFORMEDNESS’. This is known as Youden’s J satistic. This statistics can take values between 0 and 1.

Youden’s  J statistic is defined for every point on the ROC curve . The point at which Youden’s  J satistics reaches its maximum for a given ROC curve can be used to guide the selection of the threshold to be used for that classifier.

I hope this post does the job of providing an understanding of ROC curves  and AUC. The  Python program for simulating the example given earlier can be found here .

Please feel free to adjust the mean of the distributions and see the changes in the plot.