Brain-Score

To what extent do Artificial Intelligence models mirror the functioning of the mammalian brain? Research at this intersection of AI and neuroscience could answer many questions about how the brain's mechanisms for intelligence were shaped and provide us with digital twin models. The Brain-Score platform is an open-source, community-driven project that compares machine learning models to the biological activity measured experimentally in the brain. Brain-Score began as Martin Schrimpf’s doctoral project in MIT’s Department of Brain and Cognitive Sciences, and it has been further developed in a collaboration between the Quest and EPFL, where Dr. Schrimpf is now a faculty member.

Researchers upload AI models for vision and language processing to the platform and see how closely the models resemble neural activity in the brain as well as human behavioral choices, across dozens of experimental datasets. Conversely, researchers can use the latest models to prototype new ideas and experiments — and can test the alignment of potential models. Recent updates to the platform allow greater participation and collaboration: A modular plugin system simplifies the integration of data, metrics, benchmarks, and models, allowing researchers to contribute more easily to the project. 

Brain-Score 2.0 has expanded its scope to include the human language system with the introduction of Brain-Score Language. This addition underlines the platform’s domain-agnostic approach to model benchmarking and paves the way for the evaluation of large language models (LLMs), such as those used in industry.