The era of Artificial Intelligence: Part 1 – Neural networks & Deep learning

Artificial Intelligence has been discussed for a long time and often inappropriately: some say that it represents a great opportunity, while others say it could be the biggest threat to humanity. But many do not even know what it is exactly. Perhaps because it is a discipline by poorly defined contours, debated among scientists and philosophers, which presents theoretical and practical aspects as well as ethical.

We can generally define the AI as a computer’s ability to perform typical functions and reasoning of the human mind: according to this definition it is clear that today we already are surrounded by examples of AI, for example Google. The search engine of Mountain View, in fact, no longer shows just lists of links, but also direct answers to various questions (try typing “obama birth date” or “David Beckham’s wife” and see the results). To do so, DeepMind, the AI division of Google, uses sophisticated algorithms that help to understand the context of the questions.

These algorithms make sure that the system learns itself to “reason” using the so-called deep learning technique. One of the first tests an AI software has to resolve is learning the rules of a videogame: this is a “classic” test in deep learning processes because videogames represent a world of information among which the software needs to find consistent patterns. The idea is to reproduce in the machine the way the neurons work in the human brain. The programmers don’t write every rule of the “world” where the program is located, but leave it to the latter to derive some schemes from the large amount of information it receives. The concept has been applied to information technology for a long time, but only recently is bringing interesting results.

DeepMind researchers, in fact, have found a way to make an AI play some videogames without previously teaching any rules, forcing it to learn from its mistakes and the low scores obtained in the previous matches. This led to the creation of software capable of playing and winning in games like Breakout, the ping-pong simulator and even Go, a popular asian board game. Two months ago AlphaGo, a software created by Google DeepMind, unexpectedly defeated the South Korean Go champion Lee Se-dol. This result was widely recognized for the impact it has had on artificial intelligence research because the level of complexity of Go is much higher than that of chess or other games and the calculations of the possible winning moves would require a huge amount of time and computing power. AlphaGo managed to defeat the man to Go because it has learned not only to play Go but also to imitate the way a Go player thinks, adding a level of uncertainty and instinct to its gaming skills.

To be continued…