“Comparing Neural Representations in AI Systems and the Human Brain “

Abstract: This talk examines the intersection of artificial and biological intelligence by comparing neural representations in AI systems and the human brain. Leveraging multimodal vision models, such as LLaVA and GPT-4o, we explore how these systems emulate cognitive and visual perceptual processes, including emotion detection and spatial reasoning. First, we demonstrate that current models are closely aligned with neural computations in the human ventral visual stream, which is critical for object recognition. For tasks engaging this system, AI models exhibit ‘superhuman’ performance, often meeting or surpassing human capabilities. However, we show that for visual processes requiring computations performed by areas outside the ventral stream, such as spatial transformations, current AI models fail. This misalignment leads to a significant breakdown in performance, even for basic spatial reasoning tasks. Using voxel-wise encoding models, we map the computational alignments and gaps between AI models and human neural systems. Our findings offer insights into the extent of AI’s alignment with human brain functions and outline potential directions for enhancing AI’s capacity to handle more complex cognitive and perceptual tasks.

Bio: Bridget is a post-baccalaureate research scientist in the Murray lab in the Psychology department at UW. Her interest lies in the areas of computational neuroscience and NeuroAI, where she is currently comparing human cognition to AI systems. Her research aims to understand and model the neural processes within AI systems to contribute to model interpretability while building better models of the human brain.

The 2024-2025 seminars will be held in person, and are free and open to the public.