April 5, 2022 – Artificial intelligence systems are being built to help diagnose diseases, but before we can rely on them with the responsibilities of life and death, AI must develop a personality: admit mistakes.
And the truth is: they can’t do that… but.
Today, AI can often provide the most accurate answer to a problem before it can be realized that it is wrong, according to researchers from the University of Cambridge and the University of Oslo.
This serious shortcoming, they show, is rooted in a mathematical problem.
Some mathematical words cannot be proved true or false. For example, the same math that most of us learned in school cannot be used to find answers to simple and difficult questions and then prove our commitment to use it.
Maybe we gave the right answer or we didn’t, but we need to look at our performance. This is something that computer algorithms cannot do, though.
It was a mathematician first discovered by mathematicians Alan Turing and Kurt Gödel in the early 20th century.th Some mathematical problems cannot be proven.
Mathematician Stephen Smale listed this AI deficiency among the world’s 18 most common mathematical problems.
Building on the mathematical paradox, the researchers, led by Matthew Colbrook, PhD, from Cambridge University’s Department of Applied Mathematics and Theoretical Physics, devised a new way to isolate the problem areas of AI.
To the Proceedings of the National Academy of Sciencesthe researchers documented conditions where AI neural networks – which are modeled based on a network of neurons of the human brain – could be actually trained to produce reliable results.
The first step is to make AI systems more intelligent.