Technology

Opinion – Marcelo Viana: Polanyi’s paradox and machine learning

by

In the 1966 book “Tacit Dimension”, the Hungarian-British philosopher Michael Polanyi (1891–1976) pointed out that we perform numerous tasks without being able to express how we do it. “We know more than we can say,” he summarized.

This would constitute a serious obstacle to the idea of ​​artificial intelligence. Computer programs consist, so it was thought at the time, of a set of instructions that completely and accurately describe what is to be done. If we don’t know how to explain how we recognize a face or choose a move in chess, how can we write code explaining to a computer how to perform these tasks? There would be an intrinsic superiority of human intelligence over artificial intelligence: its ability to do things it cannot describe.

However these are some of the many tasks in which artificial intelligence has made spectacular progress in recent years, since the advent of machine learning methods. We discovered how computers can learn to perform complex tasks based on examples, on real data, without us having to explain exactly what they should do.

The Chinese game go has very simple rules, and yet it is one of the most complex there are, due to the astronomical number of configurations that can occur (much more than in chess). The Alpha Go Zero computer program used “deep learning” techniques to learn to play in just a few hours, competing with itself without human intervention. Today, he is the undisputed world champion of go, chess and other games, with a performance no human can rival.

Artificial intelligence skeptics do not fail to point out that in other areas, such as the development of autonomous vehicles or domestic robots, progress has been much slower. In my view, this just reflects the fact that these are even more complex problems, with many more variables than a game board, and it’s only a matter of time.

Algorithms already make reliable diagnoses in certain areas of medicine. They already compose compelling songs in the style of Bach. The ultimate challenge will be to prove theorems in the style of Gauss and Euler!

artificial intelligencedeep learningleafmachine learning

You May Also Like

Recommended for you