As with all the things I do not understand, I spend some time trying to understand them. And this is what I understand about Artificial Intelligence.
What Artificial Intelligence is
Artificial Intelligence – using AI for the sake of simplification – is an evolution of information technology capable of imitating human intelligence. Now, since words are important, I consulted Treccani (the Italian-language encyclopedia), and the word “intelligence” has this meaning: “Complex of psychic and mental faculties that allow us to think, understand or explain facts or actions, to elaborate abstract models”
Meanwhile, the “Artificial Intelligence” coupling is explained as: “Discipline that studies if and how the most complex mental processes can be reproduced through use of a computer. This research develops along two complementary paths: on the one hand, Artificial Intelligence tries to bring computer functioning closer to the capabilities of human intelligence; on the other, it uses computer simulations to make assumptions about the mechanisms used by the human mind”.
Hence two sides: “try to bring computer functioning closer to the capabilities of human intelligence” and “use computer simulations to make assumptions about the mechanisms used by the human mind”.
Intelligence has nothing to do with consciousness, this must be clear.
Information technology is providing us with the ability to assemble important amounts of data to enable us to see a trend, a tendency, a symmetry that we might miss.
AI like Lego?
When I think of AI, I think of the box of Lego I had when I was a child. I opened the box and there were hundreds of little bricks, all different, which in theory could have given life to thousands of different forms; in practice, I could only do what I could and could imagine.
So, from my point of view, here lies all the strength and all the limitations of this model.
What I can do
I can assemble data, but in doing so I will always use models known to me. I will always try in some way to understand how to assemble them and, somehow, I will tend to force a result.
If I can build houses with Lego, I will build houses. Big, small, multi-story, with trees or without … but always houses.
If I have to imagine a different model, I will have to force myself to use it, try it and try it again until I find a way to put it into practice. Because I am not only intelligent, I am conscious: that is, I will always tend to mix what I can do or can learn with my emotional experience.
What AI can do
I like to think that the true strength of AI is the lack of consciousness and partly of structure. That is, from my point of view, Artificial Intelligence is stupid in the healthiest sense of the term. When we take data and assemble them, AI processes millions of answers, sometimes far from reality; other times, however, having no constraints but pure processing capacity, it finds trends, highlights scenarios and gives answers that I, with my previous experience, could not even make assumptions or know intuitively.
Returning to my box of Lego, I build not only houses, I build airports, clouds, monsters, meadows with flowers, and so on and so forth. Sometimes I do more: I use all the pieces, even the strange ones, the ones I have never used, because – in fact – I never understood what they were for. And thus, in a way, this teaches me.
What it teaches me
The ability to process a quantity of data that is impossible for me; it teaches me not to leave anything out.
How it does it
The techniques most used (and forgive the simplification) for the processing of algorithms are:
Decision trees, or decision-making systems based on logical trees where each node is a conditional function. Each node verifies a condition (test) and associates it with a property of the environment (variable). For example, I could use an algorithm of this type to understand what the best day to go running is. The algorithm will do something like this: Monday: it’s raining/the sun’s out/it’s cloudy – if we go into the sun’s out node, wind/humidity/temperature and so on will be analyzed; the answers are in the leaves on my tree.
Bayesian classifiers – as the name implies – are algorithms linked to Bayes’ Theorem which, in practice, serves to calculate the probabilities that a certain thing will occur, knowing the frequency with which that thing has occurred previously in a similar situation.
Analysis of the main components, or according to Wikipedia, “the primary purpose of this technique is the reduction of a more or less high number of variables (representing as many features of the phenomenon analyzed) in some latent variables (feature reduction)”. In practice, the complexity of the data to be analyzed is reduced, taking only those with greater variance.
Kernel machines (also called support vector machines), which aim to solve the problem of learning starting from a training set of data whose characteristic parameters are known. A typical example of application is the recognition of handwritten figures, for example postal codes written on envelopes. The problem to be solved, which is to recognize the identity of each figure starting from a matrix, is a problem of classification, in which the aim is to assign each symbol to one of ten possible categories: (0, 1, … 9).
Ensemble methods, that is, algorithms which include previous classifications (Bayesian, decision-making, …) and analyze the final estimates to obtain more effective forecasts.
This overview (certainly not exhaustive and absolutely simplified by my ignorance on the subject) highlights an important element: deciding which algorithm to use will determine, in part, the type of response we will obtain. That is, our ability to look beyond is still an element of great importance.
This is to say that it is we who are still more intelligent.