ERC Starting Grant for Eric Schulz to investigate self-learning AI systems

Research program aims to examine self-motivated learning of Artificial Intelligence
 

Max Planck Research Group Leader Eric Schulz at the Max Planck Institute for Biological Cybernetics has won a 1.5 million euro grant from the European Research Council (ERC) to investigate learning in large language models such as ChatGPT.
 

Can algorithms be developed in the near future that will acquire new knowledge on their own initiative, without being trained by specialists to do so? And what do inducements of self-motivated learning look like? Eric Schulz's research group at the Max Planck Institute for Biological Cybernetics is using methods from cognitive psychology to better understand the processing of artificial intelligence and find answers to these questions.

"With the ERC Starting Grant we have acquired, we want to focus on the increasingly powerful basic models of chatbots.  Using experimental methods from the psychological literature, we plan to identify and systematize the underlying mechanisms of learning processes of artificial intelligent systems. Using computational methods, we aim to assess decision-making and reasoning capabilities in both the pure language domain and the combined language and vision domain, and how they evolve from generation to generation. We also want to see whether prompts can be standardized to the extent that they lead to appropriate subsequent behavior," comments Eric Schulz, Max Planck Research Group Leader at the Max Planck Institute for Biological Cybernetics in Tübingen.

Eric Schulz is also convinced that artificial intelligence will completely change our working and private lives in just a few years: That, for example, a smartphone will allow fluent conversations comparable to a human counterpart, or that on the one hand annoying, but on the other hand also very creative tasks will be taken over by AI. But how these tasks are learned and optimized independently by AI is what Eric Schulz's ERC Starting Grant aims to clarify in the coming years.

It is also clear to Schulz that the current set of scientific theories and methods of cognitive science for the study of artificial systems must adapt. He therefore proposes a new sub-discipline of cognitive science that specializes in specific perception, action and decision processes of artificial intelligent systems and that pools the knowledge of researchers from different research fields. Their common goal: to create processes based on artificial intelligence that come ever closer to human information processing.