Evolution of Deep Learning: a detailed discussion

Evolution of Deep learning

The evolution of deep learning has experienced many ups and downs since the last few decades. One time it rose to the pick of the popularity, expectations were high and suddenly some setbacks in experimental trials created a loss of confidence and disappointments. This article will cover this journey of deep learning neural network from its inception to its recent overwhelming popularity.

Background of Machine learning

This all started with very basic concepts of probabilistic modelling. These are very elementary statistical concepts from the school syllabus. This was the time even before the invention of the term machine learning. All models, functions were solely crafted by the human mind. 

Probabilistic models

These models are the first step towards the evolution of deep learning. These models are developed keeping in mind real-world problems. Variables having relationships between them. Combinations of dependent and independent variables were used as inputs to these functions. These models are based on extensive mathematical theory and more empirical than practical.

Some popular such probabilistic models are as below:

Naive-Bayes classification

It is basically the Bayes theorem with some naive assumption hence the name. The concept of this modelling was established long back during the 18th century. The assumption here is “all the features in the input data are all independent”. 

For example, suppose a data set has the data of some persons with or without diabetes disease and their corresponding sex, height and age. Now the Naive Bayes will assume that there is no correlation between all the features between sex, height and age and they contribute independently towards the disease. This assumption is called class conditional independence.

So how this assumption helps us to calculate the probability? Suppose, there is a hypothesis H which can be true or false and this hypothesis gets affected by an event e. We are interested to calculate the probability of the hypothesis being true given that the event is observed. So, we need to calculate: P(H|e)

According to Naive Bayes’ theorem

Here, P(H|e) is called the posterior probability of the hypothesis with the information of event e and can not be easily computed. So, we need to break down it as in equation 1. Now we can calculate each of the probabilities separately from the frequency table and calculate the posterior probability. 

You can read the whole process of calculation here.

P(H) is the prior probability of the hypothesis before observing the event.

Logistic regression

This regression modelling technique is so basic and popular for almost all classification problems that it can be considered as the “Hello World” of Machine Learning. Yes, you have read it right. It is a process for classification problems. Don’t let the word regression in the name misguide you. 

It is originally a regression process which becomes a classification process when the process involves a decision threshold for the prediction. Deciding a threshold for the classification process is very important and tricky one too.

We need to decide the decision threshold depending on the particular case in hand. There can be four types of responses in case of classification problems which are “true positive”, “true negative”, “false positive” and “false negative” (read details about them here). We have to fix the probability of one type of occurrence while reducing another depending on its severity.

Example and basic concept 

For example, take the case for a severe crime and it is to decide if the person should be hanged or not. It is a problem of binary classification with two outputs guilty or not guilty. Here the true positive case is the person found guilty when he actually has committed the crime. On the other hand, the true negative is the person found guilty when he has not committed the crime.

So, no doubt the true negative case here is of very serious type and should be avoided at any cost. Hence while fixing the decision threshold, you should try to reduce the probability of true negative while fixing the probability of true positive cases.

Unlike linear regression predicting the response of a continuous variable, in logistic regression, we predict the positive outcome of a binary response variable. Unlike linear regression which follows a linear function, a logistic regression has a sigmoid function.

The equation for logistic regression:

Equation for logistic regression

Initial stages of evolution of Deep Learning

Although the theoretical model of deep learning came in 1943 by Walter Pitts, a logician and Warren McCulloch, a neuroscientist. The model was called McCulloch-Pitts neurons and still regarded as a fundamental study on deep learning.

The first evidence of the use of neural networks in some toys for children made during the 1950s.  The same year the legendary mathematician Alan Turing proposed the concept of Machine Learning and even gave hints about the genetic algorithm in his famous paper “Computing machinery and intelligence”.

Alan Turing
Alan Turing (Image source: http://rutherfordjournal.org)

In 1952, Arthur Samuel first time coined the term Machine Learning. He is known as the father of machine learning. He with his association with IBM also developed the first machine learning programme.

The perceptron: the perceiving and recognizing automaton” a research paper published in the year 1957 by Frank Rosenblatt set the foundation of Deep Learning network.

In 1965 mathematician Alexey Ivakgnenko and V.G. Lapa arguably developed the first working deep learning network. Ivakgnenko for this contribution is considered as the father of deep learning by many.

The first winter period

The period between 1974-80 is considered as the first winter period. It is a long rough period faced by AI research. A critical report submitted by Professor Sir James Lighthill on AI research as asked by UK parliament played a major role to initiate this period.

The report was very critical about the AI research in the United Kingdom and was in the opinion that nothing has been done in the name of AI research. All expectations about AI and deep learning were all hype; creation of a robot was nothing but a mirage; such comments were very disappointing and resulted in the /retraction of research funding for most of the AI research.

Invention of Backpropagation algorithm

Then in during 1980, the famous Backpropagation algorithm with Stochastic Gradient Descent (SGD) was invented for training the neural network. This can be considered as path-breaking discovery as far as deep learning is concerned.  These algorithms are still the most popular among deep learning enthusiasts. This algorithm only led to the first successful application of Neural Network.


Come in 1989 we got to see the first real-life application of Neural Net. It was Yann LeCun who made this possible through his tireless effort in Bell Labs to combine the ideas of Backpropagation and Convolutional neural network.

Yann LeCun
Yann LeCun

The network was named after LeCun as LeNet. It found its first real-world problem-solving use in identification of handwritten codes. It was so efficient in identifying the codes that United States Postal Service adopted this technology in 1990 for identifying the digits of ZIP codes on the mail envelopes. 

Yet another winter period; however brief one

In spite of the success achieved by LeNet, in the same year 1990, the advent of Support Vector Machine pushed the Neural Network almost extinction. It gained very fast popularity mainly because of its easy interpretability and state of the art performance. 

It was also a technology came out of from the famous Bell Labs. Vladimir Vapnik and Corinna Cortes pioneered its invention. They started working on it long back in 1963. It’s their continuing effort that resulted in the revolutionized Support Vector Machine of 1990.

Support Vector Machine: a new player in the field

This new modelling is mainly conceptualized on a kernel trick to calculate the decision boundary between two class of variables. Except for a few cases, it is very difficult to discriminate variables on a two-dimensional. It becomes far easier to understand in a higher-dimensional space. A hyperplane of higher dimensional space becomes a hyper line in two-dimensional space i.e. a straight line. This process of transforming the mode of representation is known as kernel trick. 

Below is an example of what I mean to say by higher dimension representation for classification.

SVM: data representation in higher dimension
SVM: data representation in higher dimension

In figure A, two classes of observations that are red and blue classes are classified using a hyperline. It is a straight forward case and the classification is easy. But consider the figure in B here a straight line can not classify the points.

As a new third axis has been introduced in figure C, we can see that the classes are now can be easily separated here. Now how it will look if the figure we again convert it to its two-dimensional version? see the figure in D.

So, a curved hyperline has now separated the classes very effectively. This is what a support vector machine does. It finds a hyperplane to classify the points and then any new point gets its class depending on which side of the hyperplane it resides.

Kernel trick

A kernel trick can be explained as a technique to maximize the margin between the hyperplane and the closest data points. It makes the process very easy by curtailing the need to calculate the new coordinates in the new representation space. The kernel function only calculates the distance between the pair of points. 

This kernel function is not something that SVM learns from the data. It is solely crafted by the human mind. The distance between the points in the original space to that of in the new representation space is mapped. And then the hyperplane is created through learning from the data. 

Pros of SVM

  • The process is very accurate for the limited amount of data and when data is scarce
  • It has a strong mathematical base and also in-depth mathematical analysis is possible in SVM
  • Interpretation is very easy
  • The popularity of this process was instant and unprecedented

It also sufferers from some weaknesses like:

  • Scalability is an issue. When the data set is vast it is not very suitable.
  • Modern-day databases with a huge amount of images with enormous information provided the recognition process is efficient. SVM is not the preferred candidate here.
  • It is a shallow method so feature engineering is not easy.

Decision tree

During 2000 another classification technique made its debau. And instantly became very popular. It even surpasses the popularity of SVM. Mainly because of its simplicity, ease of visualizing and interpretation it became so popular. It also uses an algorithm which consumes very limited resource. So, a low configuration of the computing system is not a constrain for the application of the decision tree. Its some other benefits are:

  • The decision tree has a great advantage of being capable of handling both numerical and categorical variables. Many other modelling techniques can handle only one kind of variable.
  • Requires no data processing which saves a lot of user’s time.
  • The assumptions are not too rigid and model can slightly deviate from them.
  • The decision tree model validation uses statistical tests and the reliability is easy to establish.
  • As it is a white box model, so the logic behind it is visible to us and we can easily interpret the result unlike the black-box model like an artificial neural network.

But it does suffer from some limitations. Like it has a problem of overfitting. Which means that the performance with training data does not reflect when an independent data set is used for prediction. It is quick to produce a result which is often lacking satisfactory accuracy.

However, since its inception in 2000, it continued its golden run till 2010.

Random forest

This technique came to improve the weaknesses of the decision tree. As decision tree was already popular for its simplicity. Random forest took no time to win the heart of all machine learning enthusiasts. 

As it overcomes the limitations of the Decision tree, it became the most practical and robust among the shallow ML algorithms. Random forest is actually ensembling of decision trees i.e. it is a collection of decision trees where each decision tree has trained with a different dataset. The more decision tree a random forest model includes, the more robust and accurate its result becomes. It is like as we consider a forest a robust one if it has many trees.

Random forest: ensemble of decision tree
Random forest: ensemble of decision tree

Random forest actually makes a final prediction from the prediction obtained from each of the decision tree models to overcome the weakness of a single decision tree model. In this sense, the random forest is a bagging type of ensemble technique. 

We can have an idea of Random forest’s popularity by the fact that in 2010 it became the most liked machine learning in the famous data science competition website Kaggle. 

The Gradient Boosting modelling was then the only other approach which came up as the closest competitor of random forest. This technique ensemble all other weak machine learning algorithms mainly decision tree. And it was quick to outperform random forest. 

In Kaggle very soon gradient boosting ensemble approach overtake random forest. And still, this technique is the most used machine learning method along with the deep learning technique in almost all Kaggle competitions.

Dark Knight rises: The neural network era starts

Although the neural network was not consistent in showing its potential since 1980. Its success when demonstrated by some researchers like from IBM etc. it surprised the whole world with intelligent machines like Deep Blue, Watson etc. 

The dedicated deep learning scientists putting their hard work in research never had any doubt about its potential and what it is capable of to do. The only constrain till then the research work was in very scattered form. 

A coordinated research effort was very much required to establish its potential beyond any doubt. The year 2010 marked the dawn of a new era when for the first time such effort was initiated by Yann LeCun of  New York University, Yoshua Bengio of the University of Montreal, Geoffrey Hinton and his group of University of Toronto and IDSIA in Switzerland.

From the group of researchers, Dan Ciresan of IDSIA first showed the world some successful applications of modern deep learning in 2011. Using his developed GPU trained deep learning network, he won some of the prestigious academic image classification competitions.  

The ImageNet

ImageNet image classification competition conceptualized by Geofrey Hinton and his group from the University of Toronto started a significant chapter in the history of Deep Learning Neural Net in the year 2012.  

Screenshot of ImageNet
Screenshot of ImageNet (http://www.image-net.org/)

In the same year, a team headed by Alex Krizhevsky and guided by Geoffrey Hinton recorded an accuracy of 83.6% in this image classification challenge. Which was on quite a hire side compare to the accuracy of 74.3% achieved by computer vision using classical approaches in the year 2011. 

The ImageNet challenge was considered to be solved when someone with a deep convolutional network (convnets) improved the image classification accuracy up to 96.4%. Since then it was the deep convolutional neural net that had always dominated the machine learning domain.

The deep convolutional neural net got recognition by the whole world after its overwhelming success. Since then all major computer conferences and programmers meet almost all machine learning solutions are based on the deep convolutional neural net.

In some other fields like natural language processing, speech recognition also the deep convolutional neural net is a dominant technology replacing other previous tools like decision tree, SVM, random forests etc. 

A good example of major players switching to deep convolutional neural net from other technologies are like the European Organization for Nuclear Research, CERN the largest particle physics laboratory in the world has ultimately switched to deep convolutional neural net to identify new particles generated from Large Hadron Collider (LHC); earlier they were using decision tree-based machine learning methods for this task.


The article presents a detailed history of how deep learning has made a long way to reach today’s popularity and use in many fields across different scientific disciplines. It was a journey with many peaks and valleys which started way back in 1980. 

Different empirical statistical methods and machine learning algorithms preceding to deep learning made way for deep learning techniques mainly because of its high accuracy with a large amount of data. 

It registered many successes and then suddenly lost in despair for not being able to meet the high expectation. It always has a true potential being more a practical technique than empirical. 

Now the question is what is there in future of deep learning? What new surprises are in stock? The answer is really tough. The history we discussed here is evidence that many of them are already here to revolutionize our life.

So the next major breakthrough may also be just around the corner or it may take still years. But the field is always evolving and full of promises of blending machines with true intelligence. After all it learns from data so it will not repeat history of failures.


  • http://www.image-net.org/
  • Chollet, F., 2018. Deep Learning mit Python und Keras: Das Praxis-Handbuch vom Entwickler der Keras-Bibliothek. MITP-Verlags GmbH & Co. KG.
  • https://www.wikipedia.org/
  • Cortes, C. and Vapnik, V., 1995. Support-vector networks. Machine learning, 20(3), pp.273-297.
  • https://www.import.io
  • Vapnik, V., 1995. Support-vector networks. Machine learning, 20, pp.273-297.
  • Schölkopf, B., Burges, C. and Vapnik, V., 1996, July. Incorporating invariances in support vector learning machines. In International Conference on Artificial Neural Networks (pp. 47-52). Springer, Berlin, Heidelberg.
  • Rosenblatt, F., 1957. The perceptron, a perceiving and recognizing automaton Project Para. Cornell Aeronautical Laboratory.

Artificial intelligence basics and background

Artificial Intelligence basics

Artificial Intelligence (AI) is a buzz word in almost all walks of our life with a meteoric growth recently. For the last few years,  it has come up as a superpower controlling the future of every scientific endeavour. Be it AI-powered self-driving car, disease detection in medical research, image/speech recognition or big data these are just the tips of the iceberg with respect to the enormous possibilities Artificial Intelligence capable of.  This article covers the Artificial intelligence basics with its genesis including modern history.

Artificial intelligence is a broad term encompassing both Machine Learning and Deep Learning. In which Machine Learning is again a bigger domain with the subdomain Deep Learning. These three domains of advance computing can be represented by the following diagram.

Artificial Intelligence basics: Machine Learning and Deep Learning as sub domains
Artificial Intelligence basics: Machine Learning and Deep Learning as sub domains

Background of Artificial Intelligence, its genesis

Before we start with the Artificial intelligence basics, we should know its background. The first instance of any machine having some intelligence akin to human was developed by Charles Babbage and English mathematician Lady Ada Lovelace of Victorian England during 1830-40.

It was called a mechanical computer and had the capacity to perform different mathematical computation. The machine algorithm she developed lead to the creation of an early computer which just existed only in paper till then. So, Ada Lovelace, the daughter of famous poet Lord Byron was named the world’s first computer programmer

Turing machine: one step towards modern computer

Another similar example is the Turing Machine developed by Allan Turing in 1950. It can be designated as the first instance of a machine having Artificial Intelligence. He wrote a famous article on Turing Machine titled “Computing Machinery and Intelligence”. 

Turing machine was the first realized model of a computer. Turing invented it while he was working at Cipher school at Bletchley Par and the mission was to break the German Enigma code during the Second World War. It was theoretically similar to modern-day electronic computers. In 1951, the US got its first commercially available electronic stored-program computer named UNIVAC.

The modern history of Artificial Intelligence

After that many years passed with lots of trials and errors, research and development without any significant advancement in the field. The main limitation was lack of training data as images are not abundant at that time and also the computing power also insufficient to analyze the voluminous data. 

However, the scenario took a sharp turn as soon as the advent of computers with higher computational power. The term Artificial Intelligence was first coined in a conference at Dartmouth College, Hanover, New Hampshire in 1956. Again a group of researchers threw themselves to unveil the superpower of AI.


Anyway, critics are always there and their argument against AI now becoming more prominent due to lack of its practical evidence. The government also appeared to be convinced with the argument due to a lack of success in any of the AI projects. As a result funding towards all AI research projects got stopped. It was a big blow and eventually, a winter period started in AI research during 1974 and lasted till 1980

In 1980 the AI research came to the headline for a brief period when the British government showed some interest with an intention to compete with the Japenese advancement in AI research. But that did not last long; soon due to measurable failure of some early-stage computers pushed the field into another prolonged winter period which lasted for long seven years (1987 to 1993).


But the winning spree of AI was just a matter of time and inevitable. As industry leaders like IBM set foot in the AI industry and took the challenge to show the world what AI is capable of, things start to change. A team of highly qualified scientists and computer programmer threw themselves in this mission and the result was pathbreaking.

Deep Blue: the chess champion supercomputer

The first big success of the AI project was the creation of the supercomputer Deep Blue by IBM. The computer created history when it defeated the then world chess champion, Garry Kasparov on May 3rd, 1997. 

Deep Blue Vs Garry Kasparov
Deep Blue Vs Garry Kasparov (Image source: CBS news, Sunday Morning)

Back then it was so surprising that the reigning champion was not ready to admit that he has lost to a computer with Artificial Intelligence. He was crying foul play and under the suspicion that it was some grandmaster actually playing for the computer.

The computer was so accurate in making the moves but without any human emotions. Where Garry lagged behind being a human. This is where a computer always steps ahead of human being, applying only hard logic based on the vast amount of information fed to it. This victory of Deep Blue over human intelligence ushered a new age of Artificial Intelligence.

Watson: the question-answering AI-based computer

Another historic foot of establishing supremacy over human intelligence achieved by AI in 2011 when a supercomputer named Watson won the famous Quiz show called Jeopardy. In this competition, Watson defeated the defending champions Ken Jennings and Brad Rutter.

Watson and the Jeopardy challenge
Watson and the Jeopardy challenge (Image Source: IBM Research)

Watson is a question-answering computer created by IBM’s DeepQA project in the year 2010 based on Natural Language Processing. Mr David Ferruci of IBM was the key brain behind the idea of Watson. And it got the name after the founder of IBM’s founder and first CEO Thomas J. Watson.

Artificial Intelligence basics

The concept of Artificial Intelligence just reversed the traditional idea of finding a solution for any data-oriented problem. The classical programming or statistical modelling approach usually set the rules first, then apply it on the input data to achieve estimation result. Whereas Artificial Intelligence uses the example answer data along with the input data to learn the rules. See the below schematic diagram to understand it:

Artificial Intelligence basics: Difference between classical programming and AI
Artificial Intelligence basics: Difference between classical programming and AI

This concept of Artificial Intelligence suggests that it gives more emphasis on the hands-on training part. To learn from the data. Indeed this process needs a large amount of data so that the algorithm can be certain about the actual relation between the variables. Thus the idea is to establish the rules more often empirically than theoretically.

The concept of Artificial Intelligence is not a new one though. The concept first came into existence long back in 1950. During its inception, besides the concept of Deep Learning and Machine Learning, it did contain some hardcore programming rule also. For example, playing a chess game back then comprised of a lot of rules programmed to the computer. Such kind of Artificial Intelligence got a name Symbolic AI.

During 1980 the concept of Expert Systems got the limelight across the industries. An expert system on any topic actually provides an interactive information delivery system. Here a machine can play an expert role and based on the user’s input provides suitable information. In the process of developing such expert systems, the Symbolic AI transformed into Machine Learning.

Components of Artificial Intelligence

This has three main components as shown in the above figure:

Input data:

This is very obvious and also common in traditional programming or statistical modelling. We need to feed the input data in order to arrive with the estimation. The sample data in our hand either labelled or non-labelled plays this role as input data. 

Labelled data:

This is the unique part in case of Artificial Intelligence. We need to provide some example answer data to train the programme. The larger the example answer data, the more accurate is the training. This example data set is the labelled data here. As both the variables feature and label are present here. We expect the algorithm will learn from this example and identify the relationship between them.

Error optimization:

This is the third important component which calibrates the algorithm identifying how close is the estimation to the actual value. There are several metrics which provide a good measure of how good the model is performing.

Algorithm to represent the input data

In nutshell, this is the main essence of Artificial Intelligence. All machine learning or deep learning algorithms try to find out some effective way to represent the input data. This representation is of utmost importance as this is the key for successful prediction. 

For example when the problem in hand is to identify any image and the image has colour composition Red, Green and blue; then a very effective way to represent the image can be to identify the number of pixels with red colour. In similar fashion in case of speech recognition, if the algorithm can represent the language and voice modulation effectively the accuracy of recognition gets much higher.

An example of data representation

Here is an example of this representation problem with an easy graphical classification problem. This example I have read in the book “Deep learning with Python” by Keras creator and Google AI researcher Francois Chollet. It is a great book to start your journey with Artificial Intelligence.

Separating the different colour dots using data transformation
Separating the different colour dots using data transformation

See in the above figure the scattered points with two colour groups red and blue. The problem is to find out some rule to classify these two groups. A good solution to this representation problem is to create a new coordinate like the below figure. Now after the change in coordinates, the different colour dots can be easily classified with a simple rule which is the dots are blue when X>0 and red when X<0.

Data representation changing coordinates
Data representation changing coordinates

AI algorithms: not creative but effective 

This types of transformations are handled by Artificial Intelligence algorithms automatically. Like this coordinate change, other transformations like linear transformation, nonlinear operations, etc. all frequently used functions and are available for Artificial Intelligence algorithms to choose from a predefined space called Hypothetical space. In this sense Artificial Intelligence algorithms are not very creative, all they do is to select functions from this space of possibilities.

Although the algorithm is not creative, often does the work. The algorithm takes the input data; then applies suitable transformation from the Hypothetical space; the algorithm takes the help of the feedback signal obtained from the output and expected output and with this guidance, attempt to represent the input data.

The following diagram represents the flow of information process for ease of understanding.

Artificial Intelligence basics: Schematic of AI algorithm functioning
Artificial Intelligence basics: Schematic of AI algorithm functioning

Final words

So, in the simplest terms, Artificial Intelligence is all about learning through trials and examples. You provide lots and lots of example answers and the algorithm will go on perfecting itself. Unlike other prediction algorithms which reaches a plateau after a certain number of trails, AI algorithms keep improving.

A good practical example such learning process is Google’s Quick Draw. It is an AI-driven drawing game hosted by Google. As claimed by Google, it is the world’s largest doodling data set and you can also your drawing sample to it.

A screenshot of Google’s Quick Draw
A screenshot of Google’s Quick Draw

It is an experimental research on the use of AI. You will surprise to see how effortless and quick the drawing it offers using AI. You can draw a picture in less than 20 seconds time! And the reason behind its so high accuracy in pattern recognition is again as I mentioned, a huge database of example answers. Almost 15 million people have uploaded more than 50 million drawings in the database.  

Not only drawing it is the collection of several other experiments with music, video, natural language processing and many more with open access code. You can try the codes as they are open-sourced and also add your own code of AI application.

Expectations from AI should be rational and for Long term

One problem with Artificial Intelligence was the possibilities were always hyped out of proportions. The goals and expectations were set for a too short term. The obvious result of which was disappointment and loss in interest. Such disappointment resulted in two winter period in AI research as I have mentioned before.

Such winter periods slow down the development process for years together and not at all good for the researchers and scientists putting tremendous effort in AI research. They become the victims of the irrational hype created by press and media and some over enthusiasts.

When the dreams got shattered all research projects experience a crunch in research funding. The scientists who may be at the verge of some significant result got stuck with their research just because of insufficient fund. This is very heartbreaking and may deprive a scientist of his life long research achievements.

Many of the expectations from AI technology during 1960-70 are still far-reaching possibilities even in 2020. Similarly, the hype with AI in recent years may be an exaggeration too and may lead to another winter period.


So, we need to be very cautious in making realistic expectations out of AI. Instead of setting short term goals, we should look for a long term broad objective. Should give the researchers sufficient time to proceed with their research and development activities.

There is no denying that AI is going to be our everyday best friend. It is going to make our lives much much easier in the coming days. The day is not very far when we will take help of AI in every problem we face, we will take suggestion when we will feel sick, it will help to educate our kids, take us to our destination, help us to understand a foreign language and in doing so AI will take the whole humanity to a newer level of evolution.

This is not an unrealistic expectation and the day will eventually come. We just need to keep patience and have faith on highly talented AI scientists working hard to make this dream a real one.