Using Siri or any voice assistant these days has become part of our everyday life to perform a variety of tasks quickly and easily. However, not everyone knows that when we exploit the potential of these tools, we are actually approaching deep learning. A term, this, rendered in our language as deep learning, some call it hierarchical learning, and that is linked in two ways to the concept of machine learning and, more generally, to that of artificial intelligence, or AI, from the English Artificial Intelligence. If machine learning is a specific branch of artificial intelligence, in the same way the technology under examination in this guide is a subset of the former, and it implies something much broader than simple multi-level machine learning. Consequently, as we already mentioned, the applications are innumerable, and we will naturally go together to discover the main ones in the next paragraphs.
The definition of deep learning
The technical definition of deep learning or deep learning is that of a class of machine learning algorithms, i.e. the research field of machine learning, the aforementioned machine learning, and artificial intelligence that is based on different levels of representation, corresponding to hierarchies of characteristics of factors or concepts, where high-level concepts are defined on the basis of low-level ones. The definition of the Artificial Intelligence Observatory of the Politecnico di Milano is even more comprehensible to non-experts: deep learning is in fact described as a set of techniques based on artificial neural networks organized in different layers, where each layer calculates the values for the next one so that the information is processed in an increasingly complete way.
It is no coincidence then that among the deep learning architectures are usually mentioned deep neural networks, convolution of deep neural networks, deep belief networks, and recursive neural networks. All of them represent an approach according to which learning takes place thanks to data obtained through algorithms, mainly of statistical computation.
You must therefore understand that the large amount of data processed by neural networks performs a "path" very similar to what happens in the human brain, which inspires the operation of the same artificial networks. Many are the researchers and scientists known for their commitment to deep learning, such as Andrew Yan-Tak Ng, among others founder of Google Brain, Ian J. Goodfellow, recognized as one of the best innovators in the world under 35 by MIT Boston, Yoshua Bengio, Ilya Sutskever, and Geoffrey Everest Hinton, one of the key figures in artificial intelligence.
It is their contributions that now allow us to define deep learning as a system that leverages a class of machine learning algorithms that, first of all, use several cascading levels of nonlinear units to perform feature extraction and transformation tasks, with each successive level using the output of the previous level as input. The algorithms, then, rely on so-called unsupervised learning of multiple hierarchical levels of data features, creating a hierarchical representation. Not only that, the way they are designed, they learn multiple levels of representation that correspond to different levels of abstraction, consequently generating a hierarchy of concepts.
It wasn't until around 2000 that people started talking about deep learning. In a relatively short time, however, its uses have multiplied like wildfire thanks to technological advances and increasingly sophisticated neural networks. The first studies on multilayer neural networks were produced and published by the Japanese scientist Kunihiko Fukushima, who developed in 1975 the cognitron model, followed by the neo-cognitron model. The same scholar, introduced the idea of connection area for neurons that developed into convolutional neural networks.
In the 80s, the analysis of multilayer artificial neural networks continued in a more decisive way, but only in the last decade, especially thanks to the advent of Big Data and the overcoming of certain limitations, they are showing their full potential in a wide range of areas. Today, deep learning systems, among many other utilities, allow, for example, to identify objects in images and videos, to transcribe speech into text, or to identify and interpret the interests of online users, returning search results closer to their specific needs.
How deep learning works
As explained in the previous paragraph of the guide, deep learning bases its entire operation on the classification and subsequent selection of the most relevant data to reach a conclusion as optimal as possible. An operation that traces that of our biological brain, either to formulate the correct answer to a question, or to reach the resolution of a specific problem, or even to deduce a logical hypothesis. Arriving even, in many cases, to exceed the performance of the same humans, some of you will remember the famous case of AlphaGo, a software that in 2016 beat the world champion of Go, several years ahead of schedule.
Deep learning thus behaves in the same way as human reasoning, but using the artificial neural networks of which we were speaking, that is, mathematical-computational models based on the functioning of biological neural networks, in turn consisting of interconnections of information. As a matter of fact, a neural network is an adaptive system: it can modify its structure, made of nodes and relative interconnections, basing itself on external data as well as on internal ones, that connect and cross the neural network during the learning and reasoning phase.
Learning is then both automatic and "deep", where deep means on more levels. This kind of learning has proven to be decidedly more powerful than previous AI technologies, so much so that it has deserved unprecedented media attention in the last period. As well as, of course, scientific and economic attention. There is no shortage of limitations, but deep learning can certainly be noted for the quality of the results obtained, with a corresponding enormous advantage in learning to solve complex problems of pattern recognition. Although the demand for enormous computational capacity can be a limitation, the scalability of deep learning as the available data and algorithms increase is what differentiates it from machine learning.
The former improves its performance as the data increases, while the latter once it reaches a certain level of performance is no longer able to refine its performance. This is because the complex neural network learns autonomously how to analyze raw data and how to perform a certain task. The computer is thus able to "learn on its own", without human instructions, after an initial training phase. The ultimate goal is to save time and resources, especially in routine work, which is performed much more efficiently and quickly than any human, without any kind of effort and maintaining a virtually constant level of quality.
The applications of deep learning
In this area, giant strides have been made, although even today some of the decisions in deep learning are not fully understandable from an exquisitely human point of view. This does not prevent, as is now clear, to constantly improve the technologies related to deep learning, more thanks to the amount of data available and the availability of ultra high performance infrastructures, the reference, in particular, is to CPU and GPU. It should come as no surprise that deep learning is currently being applied in various industries, and especially that it will continue to be applied and expand into many other areas of our daily lives in the near future. Let's try to think, just to mention a few concrete cases, about driverless cars, drone robots for parcel delivery, voice and language recognition and synthesis for chatbots and service robots, or facial recognition for security issues.
There are also medical applications in radiology to detect some forms of cancer early, or the possibility to easily identify the genetic sequences of some diseases to produce more effective drugs. We can also mention automatic coloring of black and white images, simultaneous translation, classification of objects in a photograph, automatic generation of handwriting and text, along with intuitive division into captions. Similarly, automatic gaming has also been developed, with the system learning independently how to play a given game. To close, not to be underestimated are the capabilities of deep learning in exposing irregularities in system activities thanks to their independent and continuous learning, especially for dangerous cyber-attacks or "smart" video footage installed in the most advanced airports.
Having said that, it is however important to underline that deep learning cannot and should not be the best technological solution for every problem. Many researchers, especially in the last five years, are convinced that more effective, and perhaps less expensive, approaches will soon be identified that will make hierarchical cut-through learning based on the workings of the human brain obsolete . Almost as if it were a passing phenomenon, downsized to represent only one of the many and incredible! - manifestations of artificial intelligence.