2025-02-27 : CONECT seminar by Victor Boutin

• When: February 27th 14:00 to 16:00
• Where: Institut de Neurosciences de la Timone, Marseille, France.

During this CONECT seminar, Dr Victor Boutin will present his recent work in machine learning: Reverse Engineering Human Generalization using Artificial Intelligence.

Abstract: Although nearly every pair of objects we encounter is unique, we can consistently infer their properties based on knowledge acquired from previous experiences. This ability to transfer knowledge to new situations is called generalization and is pervasive in cognitive science. How do we generalize? In this presentation, I summarize my current results, and outline future perspectives, offering valuable insight to address this question. Cognitive scientists suggest a generalization circuit in the brain that i) extracts a disentangled and versatile representation of the sensory stimuli and ii) learns a powerful generative model of its environment. In my work, I use state-of-the-art deep learning algorithms to model those two aspects of brain computation. These models are validated through human/machine comparison on tasks inspired by cognitive science. As a case in point, I recently compared the generalization abilities of modern generative AI systems against those of humans on the one-shot drawing tasks. I demonstrate that the gap between humans and machines has started to close since the introduction of diffusion models, but that qualitative differences remain. Those differences are explainable by discrepancies in visual strategies between humans and current AI systems. The unique scientific method of my work not only allows me to uncover the computational mechanisms underpinning human generalization but also provides me with a principled path to create AI systems that are better aligned with human behavior.