The strictly literal translation of this term, clearly of Anglo-Saxon origin, is in-depth learning. And this is precisely the core of its meaning, because deep learning, a sub-category of machine learning, does nothing but create learning models on multiple levels. The concept is very simple. Let's imagine we are exposing a notion. We learn it and soon after we expose another. Our brain collects the input of the first and processes it together with the second, transforming and abstracting it more and more. Scientifically, it is correct to define the action of deep learning as the learning of data that is not provided by man, but they are learned due to the use of statistical calculation algorithms. These algorithms have one purpose: to understand how the human brain works and how it can interpret images and language. The learning thus achieved has the shape of a pyramid: the higher concepts are learned starting from, the lower levels. In this guide, you will learn what deep learning is and how does it work.
Deep learning has made great strides, achieving results that, until a few decades ago, were pure utopia. This success is due to the numerous achievements in the IT field, mainly related to the hardware sphere. We have seen how important it is to bring the computer to experience an increasing amount of sensitive data and how, until recently, the time to obtain such training was quite high. Today, due to the introduction of GPUs, or new units that contribute to data processing, this process has become much more streamlined. Another important aid came from the ease of finding numerous data collections (datasets) essential to train the system. Deep learning does a fundamental thing: it gives us the representation of data, but it does so at a hierarchical level and above all at different levels, managing to process and transform them. This transformation is amazing because it allows us to witness a machine that is able to classify the incoming (input) and outgoing (output) data, highlighting the important ones for the purpose of solving the problem and discarding those that are not needed. The revolution brought about by deep learning is all in the human-like ability to process data; one's knowledge at levels that are not at all linear. Due to this faculty, the machine learns and improves more complex functions ever.
In 1958 the psychologist Rosemblatt introduced the famous Perceptron to the academic world, an artificial neural network, powered by a computer that was the same size as a room. An event that aroused the enthusiasm of the media, so much so that a New York Times reporter wrote that humanity would soon witness the birth of personal computers capable of walking, talking and even being self-conscious. Perceptron, in fact, he had successfully carried out the task entrusted to him: to distinguish, after 50 attempts, the tiles marked on the right from those on the left. Due to the limitations of the single layer, Perceptron stopped, however encouraging the research to take further steps forward. It was the scientist Minsky who, with his theories and rather rude ways, extinguished the initial enthusiasm. Rosemblatt's dream revived in the early 1980s thanks to Hinton and LeCun, who published a study through which a path was proposed for teaching neural networks to correct errors. The technology of those years, however, was not sufficiently developed to allow us to see improvements. We need to get to the present day to see significant and valid developments, due also to the intervention of giants in the IT sector, which are starting to include deep learning in their commercial programs.
2012 is a pivotal date on the arduous path of deep learning. In that year, the extraordinary results obtained in an experiment conducted by Professor Hinton in Toronto are presented, during the ImageNet contest. The experiment was aimed at visual recognition, carried out using test software (benchmark), in a sector that had hitherto been difficult and never explored: the ability to distinguish 1000 visual categories. Hinton, in one fell swoop, achieved an epochal improvement: 10% in one go. In the course of this experiment, the pool of scholars presented this new technology that had allowed the machine to recognize humans and animals by comparison with millions of other images without the need for a codified intervention by man. It is a unique and important data that testify how deep learning has evolved over the years.
Imagenet returns and in 2015 causes an earthquake among supporters of deep learning. The IT giant presents the sensational results of incredible research. In their labs, they managed to create deep learning on 152 levels of abstraction. A remarkable result when compared to the 30 levels on which it was based before this research.
The concept of the level of abstraction is easy to understand. If deep learning uses only one algorithm at a time, it means that it has only one level of abstraction and so on. Microsoft managed to perform 152 different operations on the same image. From these concepts, fundamental truth is easily deduced. The more levels of abstraction, the more the machine has a great learning ability. In simple terms, it is smarter.
To understand the range of action of deep learning, it is necessary to clarify a fundamental concept, namely that relating to neural networks. Let's imagine a voice command. The same word, repeated by different people, can have nuances and inflections that change according to the individual who pronounces it. How does the PC recognize the sound, identify it? It is obvious that, from a strictly scientific point of view, a sequential algorithm cannot provide the right support.
We need to take a step back. In the age-old diatribe concerning the supremacy between weak and strong artificial intelligence, we have seen how a new and elaborate concept has crept in. A machine will never be truly intelligent if it cannot reproduce a system of reasoning that is biologically inspired by the human brain. The machine must be able to offer a valid paradigm or offer a way of "thinking”, similar to the functioning of human neurons.
What is a neuron? The neuron is the heart of the human nervous system. It is a cell that collects, conducts nerve impulses and divides into sensory, motor and intermediate neurons, i.e. related to the first two and to other neurons that contribute to data collection and transmission. There are more than 100,000 of them in our nervous system, and they are essential for receiving and transmitting signals.
The human neuron is the computational paradigm that feeds deep learning and does so through the famous Artificial Neural Networks. Understanding their meaning is very simple. A neural network tries to reproduce the functioning of the human neuron that is all those processes that occur in the brain during the learning phase and the subsequent recognition phase.
The starting point is a very simple question: is logic always necessary to understand? Let's imagine a baby of a few months. If his mother smiles at him, he responds with another smile precisely because he has understood that his mother is smiling. No one taught him, he simply learned, over time, to recognize and reproduce it. What is the factor that intervenes in this process? The pure and simple experience that drives learning and gives the brain the data it needs to understand. Experience is a constant in the development of childhood learning. The child, through practice, learns to recognize an imperative tone of voice from a sweet one, learns to perform a certain action and so on.
In “neuronal” software, this human process occurs in this way. The programmer enters known data into the machine. He knows well the result he wants to obtain and modifies the reference parameters of that neuron to obtain that precise result. What is the programmer doing? It feeds the experience into the machine in such a way that it can respond correctly even when faced with totally new data.
We have arrived at the fundamental point that explains the concept of deep learning offered by deep learning. The neural network learns through experience, reads data, building hierarchical architectures and providing advanced levels of input-output. The resounding leap forward made by research is all here: the PC is not programmed but rather trained (through supervised, unsupervised and reinforcement learning).
Artificial intelligence has had a long history since it was conceptualized in antiquity. Of course, it was only after the creation of the first computers that could be implemented in a concrete way. Different currents then developed. One of these currents was to take inspiration from the functioning of the human brain in an attempt to create artificial neurons. An artificial neuron is nothing more than a relatively simple mathematical operation. The complexity lies above all in the interconnection of several neurons. The first artificial neural network was developed in 1951 by Marvin Minsky and Dean Edmonds of Harvard University. Shortly after, in 1956, Frank Rosenblatt developed the Perceptron which aroused great enthusiasm. Scientists then relied heavily on neural networks, but the results ended up disappointing.
It is only recently, due to the advancement in computing performance of computers that the concept of Deep Learning has developed. These are neural networks with many hidden layers (i.e. many layers of neurons located between the input layers, accepting data to be processed, and the output layers, intended to deliver the result of the calculation). To the surprise of specialists, the addition of these layers of neurons had an extremely beneficial impact on the quality of the results obtained. This is what has allowed artificial intelligence to come back to the fore in recent years. Most players in the field today swear by Deep Learning. Google, Apple, Facebook, Apple and Microsoft all make their own Deep Learning library available to developers.
To understand the range of action of deep learning, it is necessary to clarify a fundamental concept, namely that relating to neural networks. Let's imagine a voice command. The same word, repeated by different people, can have nuances and inflections that change according to the individual who pronounces it. How does the PC recognize the sound, identify it? It is obvious that, from a strictly scientific point of view, a sequential algorithm cannot provide the right support.
We need to take a step back. In the age-old diatribe concerning the supremacy between weak and strong artificial intelligence, we have seen how a new and elaborate concept has crept in. A machine will never be truly intelligent if it cannot reproduce a system of reasoning that is biologically inspired by the human brain. The machine must be able to offer a valid paradigm or offer a way of "thinking”, similar to the functioning of human neurons.
The computer looks at us, observes us, scrutinize us. Is it a linguistic gamble or a utopia in pure Asimovian style? In the field of computer vision, deep learning has made great strides offering us scenarios considered science fiction purely until a few years ago. The PC fully automatically understands an image and recognizes all the elements that are part of it. From a strictly literary point of view, the computer looks and observes the surrounding world exactly like a man.
And if the concept of computer vision may still seem an abstract entity, just take a look at the reality in which we live to realize that artificial vision is part of our daily life. Twitter has the ability to recognize pornographic images, eliminating them instantly, without the need for a human supervisor. Google, in the section dedicated to photos, catalogs the images, placing them in the appropriate categories. Or Facebook, with the ability to recognize faces and tag them, demonstrates how computer vision is our reality, in which we move more or less consciously.
After understanding how deep learning works and the successes achieved over the years by scholars, the question concerning the various fields of application of this in-depth system arises spontaneously. Where and to what extent can deep learning improve our lives? A field where this methodology can give important results is undoubtedly that of medical diagnostics. The application of the concept of neural networks in this field is very simple because doctors, very often, already use algorithms, especially in the specialist field. When a doctor makes a diagnosis, he does so based on his knowledge and experience, that is, that cultural background accumulated over the years. Deep learning could successfully intervene at this point, broadening and improving the physician's knowledge. The application of deep learning successfully ranges from programs aimed at medical diagnostics to quality control in pharmaceutical manufacturing.
A fascinating field relating to the possible application of deep learning is that relating to automatic driving. We are still a long way from the commercialization of fully automatic cars, but the elaborate prototypes bode well. Automatic guidance allows you to recognize obstacles on both sides of the roadway, due to the use of sensors and cameras capable of processing images. In this case, computer vision reproduces human sight, recognizing the area in which it is moving and providing all the useful information to move safely. The push forward for the creation of automatic cars was given by the possibility, offered by deep learning, to process as many as 20 billion operations per second.
Let's imagine a large company that produces a considerable quantity of products on which it needs to carry out quality control. In small distribution, the system is simple: the human operator, through touch and sight, controls the various products. In large retailers, where the pace is high and the quantities to be controlled are enormous, deep learning can be the winning weapon. Neural networks allow artificial intelligence to be screened for quality, defects, and wrong standards and so on in a very short amount of time.
Let's take a mouse and a keyboard. Reflecting on the functions performed, we can reasonably state that they are two tools that allow, albeit in a rudimentary way, the interaction between man and machine. The future of deep learning is all in this possibility: to offer humans the ability to be understood by the machine, through the understanding of oral language and gestures. The study and development of intelligent algorithms are aimed at creating thinking machines, with which it is possible to interact without the need for a mouse and keyboard and which will serve to simplify life in all its daily aspects.
The research is moving towards the study and realization of models that can allow the machine to understand human brain processes, up to the most exciting frontier of all: the understanding of human thought and the state of mind. It is an achievement that would lead to new scenarios, especially in the medical field. Think of the wonderful opportunity to understand and diagnose difficult-to-diagnose pathologies, such as depression.
We are on the threshold of historical evolution. Startups, giants of the IT sector, pool of scientists at work: deep learning is the future and artificial intelligence the context that cannot be done without. Although there are still many steps necessary to achieve the various objectives, it can reasonably be said that in-depth learning can only improve the life of each individual. From the individual microcosm to the macrocosm of industries, intelligent automation is an essential contribution, to be managed carefully, to make human interaction with the surrounding reality easier and more intuitive.
Over 350 million photos are uploaded to the popular Zuckerberg social network every day. The analysis of these images is essential to understand the user's interests and to offer him, for advertising purposes, products or services that are in close harmony with his interests. Not surprisingly, the company that heads the social network has opened a scientific laboratory, in Paris, entirely dedicated to the development of deep learning. In-depth knowledge is also useful in the semantic field. In fact, the aim of Zuckerberg and his associates is to arrive at an immediate understanding of an inappropriate and violent language, with the immediate elimination of the suspicious post.
Not only does Facebook incentivize research and development related to deep learning and artificial intelligence. Tech giants such as Google, Yahoo or Microsoft are committed to conquering new frontiers ever and providing the user with ever more technological possibilities. The latest app of the US giant offers the possibility of detecting, among the various emails, the presence of a request or a question and alerts us. It is simple right?
A window on which we still have to work for a long time is that offered by Siri, an application of the iPhone. Siri offers the possibility to converse with a voice assistant as if it were really a person. However, it only takes a few moments to understand that in reality, beyond the telephone, we are in contact with a machine. It is a technology in its infancy, no doubt
, but no less interesting or full of surprising developments.
2830 Words
Apr 30, 2021
6 Pages
