In this article by Dipayan Dev, the author of the book Deep Learning with Hadoop, we will see a brief introduction to concept of the deep learning and deep feed-forward networks.

"By far the greatest danger of Artificial Intelligence is that people conclude too early that they understand it."

                                                                                                                 - Eliezer Yudkowsky

Ever thought, why it is often difficult to beat the computer in chess, even by the best players of the game? How Facebook is able to recognize your face among hundreds of millions photos? How your mobile phone can recognize your voice, and redirects the call to the correct person selecting from hundreds of contacts listed?

The primary goal of this book is to deal with many of those queries, and to provide detailed solutions to the readers. This book can be used for a wide range of reasons by a variety of readers, however, we wrote the book with two main target audiences in mind. One of the primary target audiences are the undergraduate or graduate university students learning about deep learning and Artificial Intelligence; the second group of readers belongs to the software engineers who already have a knowledge of Big Data, deep learning, and statistical modeling, but want to rapidly gain the knowledge of how deep learning can be used for Big Data and vice versa.

This article will mainly try to set the foundation of the readers by providing the basic concepts, terminologies, characteristics, and the major challenges of deep learning. The article will also put forward the classification of different deep network algorithms, which have been widely used by researchers in the last decade. Following are the main topics that this article will cover:

• Get started with deep learning
• Deep learning: A revolution in Artificial Intelligence
• Motivations for deep learning
• Classification of deep learning networks

Ever since the dawn of civilization, people have always dreamt of building some artificial machines or robots which can behave and work exactly like human beings. From the Greek mythological characters to the ancient Hindu epics, there are numerous such examples, which clearly suggest people's interest and inclination towards creating and having an artificial life.

During the initial computer generations, people had always wondered if the computer could ever become as intelligent as a human being! Going forward, even in medical science too, the need of automated machines became indispensable and almost unavoidable. With this need and constant research in the same field, Artificial Intelligence (AI) has turned out to be a flourishing technology with its various applications in several domains, such as image processing, video processing, and many other diagnosis tools in medical science too.

Although there are many problems that are resolved by AI systems on a daily basis, nobody knows the specific rules for how an AI system is programmed! Few of the intuitive problems are as follows:

• Google search, which does a really good job of understanding what you type or speak
• As mentioned earlier, Facebook too, is somewhat good at recognizing your face, and hence, understanding your interests

Moreover, with the integration of various other fields, for example, probability, linear algebra, statistics, machine learning, deep learning, and so on, AI has already gained a huge amount of popularity in the research field over the course of time.

One of the key reasons for he early success of AI could be because it basically dealt with fundamental problems for which the computer did not require vast amount of knowledge. For example, in 1997, IBM's Deep Blue chess-playing system was able to defeat the world champion Garry Kasparov [1]. Although this kind of achievement at that time can be considered as substantial, however, chess, being limited by only a few number of rules, it was definitely not a burdensome task to train the computer with only those number of rules! Training a system with fixed and limited number of rules is termed as hard-coded knowledge of the computer. Many Artificial Intelligence projects have undergone this hard-coded knowledge about the various aspects of the world in many traditional languages. As time progresses, this hard-coded knowledge does not seem to work with systems dealing with huge amounts of data. Moreover, the number of rules that the data were following also kept changing in a frequent manner. Therefore, most of those projects following that concept failed to stand up to the height of expectation.

The setbacks faced by this hard-coded knowledge implied that those artificial intelligent systems need some way of generalizing patterns and rules from the supplied raw data, without the need of external spoon-feeding. The proficiency of a system to do so is termed as machine learning. There are various successful machine learning implementations, which we use in our daily life. Few of the most common and important implementations are as follows:

• Spam detection: Given an e-mail in your inbox, the model can detect whether to put that e-mail in spam or in the inbox folder. A common naive Bayes model can distinguish between such e-mails.
• Credit card fraud detection: A model that can detect whether a number of transactions performed at a specific time interval are done by the original customer or not.
• One of the most popular machine learning model, given by Mor-Yosef et al. [1990], used logistic regression, which could recommend whether caesarean delivery is needed for the patient or not!

There are many such models, which have been implemented with the help of machine learning techniques:

The figure shows the example of different types of representation. Let's say we want to train the machine to detect some empty spaces in between the jelly beans. In the image on the right side, we have sparse jelly beans, and it would be easier for the AI system to determine the empty parts. However, in the image on the left side, we have extremely compact jelly beans, and hence, it will be an extremely difficult task for the machine to find the empty spaces. Images sourced from USC-SIPI image database.

A large portion of performance of the machine learning systems depends on the data fed to the system. This is called representation of the data. All the information related to the representation is called feature of the data. For example, if logistic regression is used to detect a brain tumor in a patient, the AI system will not try to diagnose the patient directly! Rather, the concerned doctor will provide the necessary input to the systems according to the common symptoms of that patient. The AI system will then match those inputs with the already received past inputs which were used to train the system. Based on the predictive analysis of the system, it will provide its decision regarding the disease. Although logistic regression can learn and decide based on the features given, it cannot influence or modify the way features are defined. For example, if that model was provided with a cesarean patient's report instead of the brain tumor patient's report, it would surely fail to predict the outcome, as the given features would never match with the trained data.

This dependency of the machine learning systems on the representation of the data is not really unknown to us! In fact, most of our computer theory performs better based on how the data is represented. For example, the quality of database is considered based on the schema design. The execution of any database query, even on a thousand of million lines of data, becomes extremely fast if the schema is indexed properly. Therefore, the dependency of data representation of the AI systems should not surprise us.

There are many such daily life examples too, where the representation of the data decides our efficiency. To locate a person from among 20 people is obviously easier than to locate the same from a crowd of 500 people. A visual representation of two different types of data representation in shown in preceding figure.

Therefore, if the AI systems are fed with the appropriate featured data, even the hardest problems could be resolved. However, collecting and feeding the desired data in the correct way to the system has been a serious impediment for the computer programmer.

There can be numerous real-time scenarios, where extracting the features could be a cumbersome task. Therefore, the way the data are represented decides the prime factors in the intelligence of the system.

Finding cats from among a group of humans and cats could be extremely complicated if the features are not appropriate. We know that cats have tails; therefore, we might like to detect the presence of tails as a prominent feature. However, given the different tail shapes and sizes, it is often difficult to describe exactly how a tail will look like in terms of pixel values! Moreover, tails could sometimes be confused with the hands of humans. Also, overlapping of some objects could omit the presence of a cat's tail, making the image even more complicated.

From all the above discussions, it can really be concluded that the success of AI systems depends, mainly, on how the data is represented. Also, various representations can ensnare and cache the different explanatory factors of all the disparities behind the data.

• Representation learning is one of the most popular and widely practiced learning approaches used to cope with these specific problems. Learning the representations of the next layer from the existing representation of data can be defined as representation learning. Ideally, all representation learning algorithms have this advantage of learning representations, which capture the underlying factors, a subset that might be applicable for each particular sub-task. A simple illustration is given in the following figure:

The figure illustrates of representation learning. The middle layers are able to discover the explanatory factors (hidden layers, in blue rectangular boxes). Some of the factors explain each task's target, whereas some explain the inputs.

However, while dealing with extracting some high-level data and features from a huge amount of raw data, which requires some sort of human-level understanding, has shown its limitations. There can be many such following examples:

• Differentiating the cry of two similar age babies.
• Identifying the image of a cat's eye in both day and night times. This becomes clumsy, because a cat's eyes glow at night unlike during daytime.

In all these preceding edge cases, representation learning does not appear to behave exceptionally, and shows deterrent behavior.

Deep learning, a sub-field of machine learning can rectify this major problem of representation learning by building multiple levels of representations or learning a hierarchy of features from a series of other simple representations and features [2] [8].

The figure shows how a deep learning system can represent the human image through identifying various combinations such as corners, contours, which can be defined in terms of edges.

The preceding figure shows an illustration of a deep learning model. It is generally a cumbersome task for the computer to decode the meaning of raw unstructured input data, as represented by this image, as a collection of different pixel values. A mapping function, which will convert the group of pixels to identify the image, is, ideally, difficult to achieve. Also, to directly train the computer for these kinds of mapping looks almost insuperable. For these types of tasks, deep learning resolves the difficulty by creating a series of subset of mappings to reach the desired output. Each subset of mapping corresponds to a different set of layer of the model. The input contains the variables that one can observe, and hence, represented in the visible layers. From the given input, we can incrementally extract the abstract features of the data. As these values are not available or visible in the given data, these layers are termed as hidden layers. In the image, from the first layer of data, the edges can easily be identified just by a comparative study of the neighboring pixels. The second hidden layer can distinguish the corners and contours from the first layer's description of the edges. From this second hidden layer, which describes the corners and contours, the third hidden layer can identify the different parts of the specific objects. Ultimately, the different objects present in the image can be distinctly detected from the third layer. Image reprinted with permission from Ian Goodfellow, Yoshua Bengio, and Aaron Courville, Deep Learning, published by The MIT Press.

Deep learning started its journey exclusively since 2006, Hinton et al. in 2006[2]; also Bengio et al. in 2007[3] initially focused on the MNIST digit classification problem. In the last few years, deep learning has seen major transitions from digits to object recognition in natural images. One of the major breakthroughs was achieved by Krizhevsky et al. in 2012 [4] using the ImageNet dataset 4.

The scope of this book is mainly limited to deep learning, so before diving into it directly, the necessary definitions of deep learning should be provided. Many researchers have defined deep learning in many ways, and hence, in the last 10 years, it has gone through many explanations too! Following are few of the widely accepted definitions:

• As noted by GitHub, deep learning is a new area of machine learning research, which has been introduced with the objective of moving machine learning closer to one of its original goals: Artificial Intelligence. Deep learning is about learning multiple levels of representation and abstraction, which help to make sense of data such as images, sound, and text.
• As recently updated by Wikipedia, deep learning is a branch of machine learning based on a set of algorithms that attempt to model high-level abstractions in the data by using a deep graph with multiple processing layers, composed of multiple linear and non-linear transformations.

As the definitions suggest, deep learning can also be considered as a special type of machine learning. Deep learning has achieved immense popularity in the field of data science with its ability to learn complex representation from various simple features. To have an in-depth grip on deep learning, we have listed out a few terminologies. The next topic of this article will help the readers lay a foundation of deep learning by providing various terminologies and important networks used for deep learning.

Getting started with deep learning

To understand the journey of deep learning in this book, one must know all the terminologies and basic concepts of machine learning. However, if the reader has already got enough insight into machine learning and related terms, they should feel free to ignore this section and jump to the next topic of this article. The readers who are enthusiastic about data science, and want to learn machine learning thoroughly, can follow Machine Learning by Tom M. Mitchell (1997) [5] and Machine Learning: a Probabilistic Perspective (2012) [6].

Image shows the scattered data points of social network analysis. Image sourced from Wikipedia.

Neural networks do not perform miracles. But if used sensibly, they can produce some amazing results.

Deep feed-forward networks

Neural networks can be recurrent as well as feed-forward. Feed-forward networks do not have any loop associated in their graph, and are arranged in a set of layers. A network with many layers is said to be a deep network. In simple words, any neural network with two or more layers (hidden) is defined as deep feed-forward network or feed-forward neural network. Figure 4 shows a generic representation of a deep feed-forward neural network.

Deep feed-forward network works on the principle that with an increase in depth, the network can also execute more sequential instructions. Instructions in sequence can offer great power, as these instructions can point to the earlier instruction.

The aim of a feed-forward network is to generalize some function f. For example, classifier y=f/(x) maps from input x to category y. A deep feed-forward network modified the mapping, y=f(x; α), and learns the value of the parameter α, which gives the most appropriate value of the function. The following figure shows a simple representation of the deep-forward network to provide the architectural difference with the traditional neural network.

Deep neural network is feed-forward network with many hidden layers:

Datasets are considered to be the building blocks of a learning process. A dataset can be defined as a collection of interrelated sets of data, which is comprised of separate entities, but which can be used as a single entity depending on the use-case. The individual data elements of a dataset are called data points. The preceding figure gives the visual representation of the following data points:

• Unlabeled data: This part of data consists of human-generated objects, which can be easily obtained from the surroundings. Some of the examples are X-rays, log file data, news articles, speech, videos, tweets, and so on.
• Labelled data: Labelled data are normalized data from a set of unlabeled data. These types of data are usually well formatted, classified, tagged, and easily understandable by human beings for further processing.

From the top-level understanding, the machine learning techniques can be classified as supervised and unsupervised learning based on how their learning process is carried out.

Unsupervised learning

In unsupervised learning algorithms, there is no desired output from the given input datasets. The system learns meaningful properties and features from its experience during the analysis of the dataset. During deep learning, the system generally tries to learn from the whole probability distribution of the data points. There are various types of unsupervised learning algorithms too, which perform clustering, which means separating the data points among clusters of similar types of data. However, with this type of learning, there is no feedback based on the final output, that is, there won't be any teacher to correct you! Figure 6 shows a basic overview of unsupervised clustering.

A real life example of an unsupervised clustering algorithm is Google News. When we open a topic under Google News, it shows us a number of hyper-links redirecting to several pages. Each of those topics can be considered as a cluster of hyper-links that point to independent links.

Supervised learning

In supervised learning, unlike unsupervised learning, there is an expected output associated with every step of the experience. The system is given a dataset, and it already knows what the desired output will look like, along with the correct relationship between the input and output of every associated layer. This type of learning is often used for classification problems. A visual representation is given in Figure 7.

Real-life examples of supervised learning are face detection, face recognition, and so on.

Although supervised and unsupervised learning look like different identities, they are often connected to each other by various means. Hence, that fine line between these two learnings is often hazy to the student fraternity.

The preceding statement can be formulated with the following mathematical expression:

The general product rule of probability states that for an n number of datasets n ε ℝk, the joint distribution can be given fragmented as follows:

The distribution signifies that the appeared unsupervised problem can be resolved by k number of supervised problems. Apart from this, the conditional probability of p (k | n), which is a supervised problem, can be solved using unsupervised learning algorithms to experience the joint distribution of p (n, k):

Although these two types are not completely separate identities, they often help to classify the machine learning and deep learning algorithms based on the operations performed. Generally speaking, cluster formation, identifying the density of a population based on similarity, and so on are termed as unsupervised learning, whereas, structured formatted output, regression, classification, and so on are recognized as supervised learning.

Semi-supervised learning

As the name suggests, in this type of learning, both labelled and unlabeled data are used during the training. It's a class of supervised learning, which uses a vast amount of unlabeled data during training.

For example, semi-supervised learning is used in Deep belief network (explained network), a type of deep network, where some layers learn the structure of the data (unsupervised), whereas one layer learns how to classify the data (supervised learning).

In semi-supervised learning, unlabeled data from p (n) and labelled data from p (n, k) are used to predict the probability of k, given the probability of n, or p (k | n):

Figure shows the impact of a large amount of unlabeled data during the semi-supervised learning technique. Art the top, it shows the decision boundary that the model puts after distinguishing the white and black circle. The figure at the bottom displays another decision boundary, which the model embraces. In that dataset, in addition to two different categories of circles, a collection of unlabeled data (grey circle) is also annexed. This type of training can be viewed as creating the cluster, and then marking those with the labelled data, which moves the decision boundary away from the high-density data region. Figure obtained from Wikipedia. Deep learning networks are all about representation of data. Therefore, semi-supervised learning is, generally, about learning a representation, whose objective function is given by the following:

l = f (n)

The objective of the equation is to determine the representation-based cluster.

The preceding figure depicts the illustration of a semi-supervised learning. Readers can refer to Chapelle et al.'s book [7] to know more about semi-supervised learning methods.

So, as we have already got a foundation of what Artificial Intelligence, machine learning, representation learning are, we can move our entire focus to elaborate on deep learning with further description.

From the previously mentioned definition of deep learning, two major characteristics of deep learning can be pointed out as follows:

• A way of experiencing unsupervised and supervised learning of the feature representation through successive knowledge from subsequent abstract layers
• A model comprising of multiple abstract stages of non-linear information processing

Summary

In this article, we have explained most of these concepts in detail, and have also classified the various algorithms of deep learning.