BrainWideBench is a benchmark built on the International Brain Laboratory Brainwide Map dataset, which includes neural and behavioral recordings from 276 brain regions across 139 mice performing a sensory-guided decision-making task. The benchmark evaluates across-animal transfer using three complementary task suites: behavior prediction, neural activity prediction, and neuron identity prediction. It systematically tests pretraining methods under finetuning and zero-shot transfer settings.
The authors found that pretraining improves performance over matched single-session baselines, but gains are heterogeneous and depend strongly on the alignment between pretraining objectives and downstream tasks. No single approach performs uniformly well across all three task suites. This result is interesting because it reveals that current methods fall short of learning general-purpose neural representations that transfer across behavior, dynamics, and anatomy, highlighting an open challenge for the field.