In The Hitchhiker’s Guide to the Galaxy, Douglas Adams talks about a supercomputer built with the aim of finding the answer to the fundamental question about life, the universe and everything. After seven and a half million years of processing, the computer comes up with the answer: 42. So another computer has to be built to understand what the fundamental question is, but that’s another story. Here what interests us is recalling the name of the computer that answered 42: Deep Thought.
This use of the adjective”deep” precedes by thirty years the name of a class of machine learning that is running up extraordinary successes, and which we are increasingly using more often, usually without knowing it: the so-called “deep learning” algorithms.
Why are deep learning algorithms important? What is involved?
Today, deep learning is undoubtedly the most powerful and versatile tool in the hands of data scientist and companies that want to add intelligent solutions to their applications, or meet needs that were previously literally in the world of science fiction.
The following can be cited as examples of applications: image recognition, particularly applied to medicine with results that surpass the eye of human experts in recognizing degenerative diseases from diagnostic images; text recognition and generation; translation of images into texts, that is to say, the recognition of text within an image and its extraction, such as an automobile license plate; recognition and understanding of audio conversations, etc.
Every day we are exposed to deep learning on social networks: when we translate text from Arabic into Chinese using Google Translator, or we recognize faces in photographs posted on Facebook (using the DeepText tool which serves this purpose).
The users of these systems are in their millions, billions: we all are. And here’s the beauty of it, that is, these techniques are all the more accurate the more data they can “ruminate”: navigating, clicking and using our now-inseparable devices all day, and often even at night, we provide a stream of unimaginable data, which makes the algorithms that use them increasingly more accurate, in a vicious and bulimic circle of data and information.
But the plethora of applications and perspectives deployed by using these techniques, which are increasingly accessible and integrated in traditional applications, goes far beyond the use that the “giants” of the IT industry make of them.
But the plethora of applications and perspectives deployed by using these techniques, which are increasingly accessible and integrated in traditional applications, goes far beyond the use that the “giants” of the IT industry make of them.
Let’s try to understand what deep learning actually is: why deep? Why learning?
It is really a question of di reti neurali, that is, simplified models of the human brain, which consist of a set of “artificial neurons”, each of which stores a number, and each of which is connected to other neurons in the network.
Typically, these neurons are placed on layers, which we can think of as the floors of a building, in which neurons are the rooms on each floor, but rooms which are a bit particular, which have no doors or windows, but only trapdoors on the floor or on the ceiling: in fact, each layer contains neurons which are not connected to each other but may be connected to the neurons of the previous layer and the next layer.
In short, in this building you can only go up and down from one floor to another but not move from room to room on the same floor.
A network with 9 neurons and 3 layers: on the input layer there are 3 neurons, on the layer interno ce ne sono 4 e sul layer di output ce ne sono 2.
Deep learning is to traditional neural networks like skyscrapers are to normal condos.
Furthermore, the algorithms that govern these deep networks are more sophisticated: in some, you can go up or down floors (recurrent networks), others are inspired by the functioning of the human eye for “scanning” data streams according to mathematical rules (convolutive networks).
They might not be able to find the answer to the fundamental question about life, the universe and everything, but these profound thoughts now pervading the IT market are a well-established and pervasive technology with which those who produce and consume software will have to increasingly reckon with.
Paolo Caressa
