The quest of IDSIA USI-SUPSI for more "human" AI

Prof. Andrea Rizzoli (image by Marian Duven, courtesy of Ticino Scienza)
Prof. Andrea Rizzoli (image by Marian Duven, courtesy of Ticino Scienza)

Institutional Communication Service

23 March 2021

As of December 2020, the Dalle Molle Institute for Artificial Intelligence (IDSIA), a joint USI-SUPSI institute, has a new director, Andrea Rizzoli, who succeeds to Luca Maria Gambardella, who led the institute for 25 years. Rizzoli, a Swiss-Italian citizen, with degree from the Milan Polytechnic and further studies in Australia, has been working at IDSIA for a long time. He is now faced with the great challenge of artificial intelligence (AI), as he recounts to science journalist Agnese Codignola in an interview with Ticino Scienza.

Neural networks and Bayesian approaches

"Most of the recent successes in AI are based on applications of neural networks, which are networks made by trying to mimic the structure of our brain, and in which artificial neurons are connected to each other and organized in layers," Professor Andrea Rizzoli explains. "Neural networks are instructed with large amounts of incoming and outgoing data and, thanks to this, they reinforce mutual connections and, with them, the ability to determine whether or not a new piece of information can fit into a learned pattern."

However, a system of this type has limitations because it cannot go further and explain how and why a certain result was produced. This is why researchers at IDSIA try to move towards networks called Bayesian, which are inspired by the theorem discovered by the Reverend Thomas Bayes in the 18th century. "With this approach - Rizzoli explains - the computer uses information about the causes of a certain phenomenon and calculates how many probabilities there are that a certain piece of information falls within what it knows. By doing so, it comes much closer to the reasoning mode typical of humans".

Successes and ethics of AI

A story that has brought both admiration and concern is that relating to DeepMind, a company co-founded by IDSIA graduate Shane Legg and sold to Google for $500 million in 2014. DeepMind has developed systems that can beat humans in complex games like Chinese Go or, more recently, help partially solve the problem of protein structure, so-called folding. [read the news item in Nature here >>]

This and other successes, on the other hand, have also raised concerns on the ethical issues related to AI, namely privacy. Recently and for the first time, participants in one of the most prestigious events in the field, the Neural Information Processing Systems (NeurIPS), held online in December 2020, were also required to reveal the ethical aspects of their research, aspects reported also in the science journal Nature. The IDSIA also addresses these issues, and in particular Alessandro Facchini, a mathematical logician, does so in order to offer a different, and crucial, look at research of a more purely engineering nature.

The "human factor" that makes the difference

The several groups at IDISA, which include about eighty researchers, perform diverse research. "Today, the IDSIA - concludes Rizzoli - no longer relies on initial funding from the Dalle Molle Foundation, but is supported by the Canton of Ticino through the university institutions and acquires competitive funding for research projects from the Swiss National Fund, Innosuisse, the European Commission and partnerships with companies. Of course, compared to the economic and human strengths deployed by our main international competitors, this is often not much. But we focus on the human factor because, as the story of Shane Legg and many other brilliant researchers who began their careers at IDSIA has shown, we believe that the most important thing is people, to work with AI."

AI and biomedical sciences

After significant work done in collaboration with IDSIA in the field of oncology, recently a new prestigious recognition has arrived from the European Union under the Horizon 2020 program, which included IDSIA in the Advanced machine learning for Innovative Drug Discovery (AIDD) project, to combine artificial intelligence systems with chemistry and pharmacology. "This is an important and innovative project," Michael Wand, senior researcher and AIDD liaison for IDSIA, explains to Ticino Scienza. - We are doing research, but there is also a teaching component, to introduce young people in a field with great potential and strong growth prospects".