Abstract
A grand ambition of machine learning is to create generative artificial intelligence capable of assisting us in our tasks. Although much progress has been made, GenAI agents still struggle to align themselves with some of the goals and thinking patterns of the human users. The mainstream solution has been to painstakingly try out different variants of the training process and scale up data and compute. But there is another way: The development of efficient, human-compatible GenAI agents can be streamlined by tapping into the methodological and theoretical advances from cognitive neuroscience.
In my talk, I outline a research program applying what is known about the human mind to address current issues in machine learning. In particular, there is a solid base of scientific facts about learning and memory which can be used to inform architectural choices. I argue that GenAI already follows some principles of human cognition and that making it even more human-like will increase its ability to cooperate with humans and solve human-made tasks.
Short BioDr. Jaroslaw R. Lelonkiewicz is an experimental psychologist specialised in language, learning and memory. In his research, he investigates how these faculties shape complex cognitive phenomena such as alignment.
Dr. Lelonkiewicz obtained a PhD from The University of Edinburgh (2017; Prof. Martin Pickering), followed by teaching and research at excellent international institutions (The University of Edinburgh, Edinburgh Napier University, Scuola Internazionale Superiore di Studi Avanzati, Monterrey Institute of Technology, Vrije Universiteit Brussles). He is now based at Universitat Pompeu Fabra. His current work is supported by a Marie Skłodowska-Curie Fellowship and a Ramon y Cajal Fellowship.
For more information on Dr. Lelonkiewicz’s profile, see ORCID or Open Science Framework.