A project to correct biases in artificial intelligence, awarded €600,000 by the BBVA Foundation News

VALENCIA, 28 Feb. (EUROPA PRESS) -

A project promoted by the University of Valencia (UV) professor Jesús Malo López to correct biases in artificial intelligence (AI) has obtained 600,000 euros from the BBVA Foundation.

This is an initiative that aims to "improve artificial neural networks by imitating natural visual neurons, a mechanism that allows us to improve some of the limitations presented by artificial intelligence," explains the academic institution in a statement.

'Taking advantage of Vision Sciences to overcome the critical limitations of artificial neural networks' is the name of the initiative, in which researchers from the CSIC and the University of Tübingen (Germany) also participate.

The professor of Optics at the University of Valencia, Jesús Malo López, is one of the promoters of this project selected by the BBVA Foundation within the Fundamentals Program for the development of fundamental and interdisciplinary research.

Together with Jesús Malo López, from the Department of Optics of the Faculty of Physics of the University of Valencia; Marcelo Bertalmío, Institute of Optics (CSIC) promotes the project; and Felix Wichmann, from the Computer Science Department of the Neural Information Processing Group at the University of Tübingen (Germany).

Artificial neural networks are computer systems that imitate the human brain and are therefore capable of processing information. They are based on biological neurons but have some limitations. According to the authors of the project "the great success of artificial neural networks is being the main engine of the meteoric rise of artificial intelligence in recent years, so it is not an exaggeration to say that they are rapidly reshaping science, industry and society." in general".

But the authors also point out that this is why it is essential to know and face the limitations that these artificial networks present. Some of these limitations are, for example, biases that can show discriminatory behavior, or lack of explainability that makes it impossible to interpret, in a way explainable to humans, how the network reaches a conclusion.

The scientific community is already applying several promising strategies to address these limitations, in ways that could alleviate or even resolve them in the future. The problem is that there are some types of limitations of artificial networks, which are called critical, for which there are no promising solutions yet and which require new advances.

The main problems come from the rigidity of conventional models and their susceptibility to attacks or tendency to have illusory perceptions. The amount of training data and the associated energy cost also stand out as a drawback. Current networks require enormous amounts of information to perform human-level work on a given task. This implies the use of very large computing resources with their associated energy cost and CO2 emissions.

Thus, the main objective of the project is to develop a framework for these artificial neural networks so that their behavior is more similar to that of human observers, in the sense that they are more resistant to attacks, easier to train and with better performance properties. generalization.

To achieve this, the team of researchers aims, on the one hand, to design new components for the networks using very recent results and techniques from Vision Sciences; on the other hand, optimize the components of the networks using key experimental results from Visual Psychophysics as training data; and finally, validate and fine-tune the new networks for fundamental problems of Computer Vision.

The project in which the University of Valencia participates is one of the five selected, from among 305 applications, by the Fundamentals Program that will receive aid of 600,000 euros each. The objective of this program is to support exploratory research into central or foundational questions of a scientific field or the intersection of several disciplines. The University's is part of the area of Mathematics, Statistics, Computer Science and Artificial Intelligence. The team has three years to develop the project.