Fri, March 1, 8:30 PM
90 MINUTES
Neuroscience-Inspired Evaluation and Architecture for Generative AI

This talk focuses on our recent research at the intersection of neuroscience and artificial intelligence, including the evaluation and enhancement of large language models (LLMs) and building gaming AI agents. I will present studies on the behavioral evaluation of various LLMs, the development of neuroscience-inspired architectures for LLMs, the investigation of latent representations and cognitive maps within these models, and the application of in-context learning to emulate human-like reasoning and decision-making. I will note challenges and solutions in auditing and improvement of LLM-based text evaluation and discuss broader implications of these advancements for creating human-like agents and generative AI. I will also discuss our practical research in gaming AI, focused on using diffusion models and behavioral methods for building human-like game agents. This synthesis of ideas from recent studies provides insights into how neuroscience-inspired approaches to AI can lead to more human-like artificial intelligence systems.

Ida Momennejad

Principal Researcher @ Microsoft Research