BrainWideBench: Benchmarking large-scale pretraining and across-animal transfer in multi-region neural recordings

TL;DR

BrainWideBench introduces a standardized benchmark for evaluating how well neural representations learned from large-scale multi-animal, multi-region recordings transfer across tasks. It assesses performance on behavior decoding, neural activity prediction, and anatomical structure recovery. Results show pretraining improves downstream performance, but no method excels across all tasks, highlighting a gap in unified representation learning.

Why It Matters

The benchmark enables consistent evaluation of general-purpose neural models across diverse tasks, accelerating progress toward scalable brain-wide representation learning. It supports reproducible research and helps identify limitations in current pretraining strategies.

Key Findings

• Pretraining consistently improves performance over single-session baselines across all task suites.
• Transfer effectiveness varies significantly based on alignment between pretraining objectives and downstream tasks.
• No single pretraining method achieves strong performance across all three evaluation dimensions: behavior decoding, neural dynamics prediction, and anatomical organization recovery.