Martin Schrimpf: Vision and Language in Brains and Machines

Vision and Language in Brains and Machines

Abstract:
While modern machine learning originated from the study of the brain and mind, it has long departed from focusing on the human implementation to intelligence. Today's AI models are thus argued by many to be inconsistent with biology. However, empirical evidence suggests that the latest models converge to surprisingly brain-like solutions. Specifically, across a battery of neural and behavioral data, we find that the models that perform best at solving ecological tasks -- such as visual object categorization and next-word prediction -- are also the models that best align with natural intelligence. We are further closing the gap to biology with neuroanatomical models that generalize as well as humans. With these models, we can guide experiments: they generate visual or linguistic inputs that predictably control neural activity, and predict the behavioral effects of neural interventions. Taken together, I will argue that we can understand the human brain and mind in engineering terms.

About the speaker:
Martin Schrimpf is an assistant professor at the EPFL Neuro-X institute, with appointments in the School of Computer and Communication Sciences ‪@ic_epfl‬ the School of Life Sciences. His research focuses on a computational understanding of brain-like intelligence. Following degrees from TUM, LMU and UNA, Martin worked at Siemens, Salesforce, and Harvard before completing his PhD at MIT. He co-founded Integreat which is now helping newcomers in every sixth city in Germany. His work has been featured in Science Magazine, MIT News, and Scientific American. He has been awarded the Neuro-Irv Open Science Prize, the Walle Nauta Award for Continuing Dedication in Teaching, and the Takeda fellowship in AI + Health.