Chatbots like ChatGPT were unthinkable a few years ago. Now they are part of our everyday life. However, we still don’t know why they work so well and how good these techniques can become in the future. With his project “beyond2”, SISSA physicist Sebastian Goldt hopes to discover the fundamental theoretical principles that allow neural networks to achieve that kind of performance and learn something about the structure of language in the process. Thanks to the European Research Council, he will have 1.5 million Euros over the next five years to pursue his dream.
Sebastian Goldt joined SISSA in 2020 as an assistant professor in the newly created Data Science group. He has now been awarded with a Starting grant from the European Research Council (ERC) worth 1.5 million euros, which will help him grow his own team. “When I first got the news, I was incredibly happy about this recognition and excited about the opportunities that come with it”, Goldt commented. “Now I’m looking forward to welcoming the new group members that I will be able to hire thanks to the ERC funding and to all the cool science we will do together.”
Goldt’s project "beyond2" will study neural networks, the key technology behind tools like facial recognition on smartphones, or chatbots like ChatGPT. These networks "learn" by processing vast amounts of data, but they work like "black boxes": “while they are effective in practice, we don't really understand how they make decisions or what patterns they pick up from data,” Goldt explained. “We will study what and how neural networks actually learn from their data, which could ultimately help improve these technologies. But that’s not all: neural networks are also a powerful model of how the brain processes information. Thus, part of the project will focus on neuroscientific questions, and in particular we will look at how interactions between individual nerve cells in the brain help it achieve its tasks.”
Read the full press release: