BrainWideBench: A Benchmark for Large-Scale Neural Pretraining and Across-Animal Transfer

Forum topic · 小凯 · 2026-09-22

Summary

BrainWideBench (arXiv:2609.22064) is a new benchmark for evaluating cross-animal transfer on multi-region neural recordings, built on the International Brain Laboratory Brainwide Map dataset. It covers neural and behavioral recordings from 139 mice performing a sensory-guided decision-making task across 276 brain regions. The benchmark is organized into three complementary task suites: downstream decoding that tests whether learned representations support behavior, masked/future neural activity prediction, and recovery of biologically meaningful anatomical organization. The authors systematically evaluate multiple pretraining approaches under transfer settings, including fine-tuning on downstream objectives and zero-shot generalization to unseen animals. Results confirm that pretraining outperforms matched single-session baselines, but transfer benefits are heterogeneous and depend strongly on alignment between pretraining objectives and downstream tasks. No single method performs consistently well across all three suites, indicating that learning representations that jointly generalize across behavior, dynamics, and anatomy remains an open challenge. BrainWideBench provides a unified, reproducible evaluation framework for measuring progress toward foundation models of the mouse brain.

Paper Overview

Field: Machine Learning
Authors: Alexandre Andre, Shivashriganesh P. Mahato, and collaborators including the International Brain Laboratory, Blake Richards, Guillaume Lajoie, Cole Hurwitz, Mehdi Azabou, Matthew R. Whiteway, Liam Paninski, and Eva L. Dyer
arXiv: 2609.22064

What BrainWideBench Is

Advances in large-scale neural recording have made it possible to collect data across many animals and distributed brain regions, raising the question of whether this scale can be exploited to learn general-purpose neural representations transferable across diverse downstream tasks. Progress toward this goal has been limited by fragmented evaluation protocols and a narrow focus on individual task domains.

BrainWideBench is a benchmark for evaluating across-animal transfer on multi-region neural recordings, built on the International Brain Laboratory (IBL) Brainwide Map dataset: neural and behavioral recordings spanning 276 brain regions from 139 mice performing a sensory-guided decision-making task.

Three Task Suites

1. Downstream decoding — evaluates whether learned representations support behavior. 2. Neural activity prediction — predicts masked or future neural activity. 3. Anatomical structure recovery — tests whether representations recover biologically meaningful anatomical organization.

Evaluation and Findings

Using the benchmark, the authors systematically evaluated multiple pretraining methods under transfer settings, including fine-tuning on downstream objectives and zero-shot generalization to unseen animals.

Key findings:

Pretraining outperforms matched single-session baselines.
Transfer benefits are heterogeneous: gains depend strongly on alignment between the pretraining objective and the downstream task.
No single method performs consistently well across all three suites; most methods were designed for only a subset of them.
Learning representations that jointly generalize across behavior, dynamics, and anatomy remains an open challenge.

Significance

By providing a unified and reproducible evaluation suite, BrainWideBench establishes a framework for measuring progress toward general-purpose models of the mouse brain.

---

*Auto-collected on 2026-09-22.*