epfl-neuroai/multimodal-brain-scaling

Multimodal Scaling Laws for Task & Data-Optimized Models of Visual Cortex

This repository hosts the result tables and accompanying metadata released with our ICML 2026 paper Multimodal Scaling Laws for Task & Data-Optimized Models of Visual Cortex.

Paper summary

Task-optimized deep networks are the leading in-silico models of sensory cortex, but progress is fragmented across datasets, modalities, and evaluation protocols, making it hard to identify which modeling choices actually drive better brain alignment. We address this gap with a unified, large-scale benchmark of model-to-brain alignment spanning 8 neural datasets (macaque electrophysiology and human fMRI / EEG / MEG) and >600 vision models evaluated under a consistent pipeline. Along three scaling axes — pretraining resources, neural fine-tuning, and the mapping from features to neural responses — we report three robust trends:

1. Pretraining saturation. Alignment improves with pretraining compute and data, but saturates at large scale across every recording modality.
2. Complementary neural fine-tuning. Hybrid task + neural objectives yield consistent, transferable gains over pure task pretraining.
3. Mapping scaling. Increasing paired stimulus--response data for the mapping yields log-linear gains that dominate alignment improvements in many regimes.

We also introduce a lightweight, subject-shared cross-attention readout that matches or surpasses per-subject linear decoders with an order of magnitude fewer parameters. Together, these results provide practical guidance for allocating data, compute, and supervision when building next-generation brain-aligned models.

Repository contents

Large result tables are stored as Parquet files under data/; small companion files (dataset dimensions, normalization constants, model metadata) live under metadata/.

Result tables (data/)

• pretraining: Pretraining-scale model-to-brain alignment results with model metadata.
• layer_search: All-layer search results used to choose model layers for benchmark evaluation.
• finetuning: Neural fine-tuning and LoRA result table.
• mapping: Mapping-data scaling and readout comparison result table.

Metadata (metadata/)

• dataset_dims.json: per-dataset stimulus counts, subjects, and ROI sizes.
• model_metadata.csv / model_metadata_compact.csv: model family, architecture, pretraining data, parameter count, and training compute for every backbone evaluated.
• nc_summary.csv: per-ROI noise ceilings used to normalize alignment scores.
• mapping_params_per_roi.csv: parameter budgets of the mapping models per ROI.

Quick start

from datasets import load_dataset

# Main pretraining-scale alignment results with model metadata.
ds = load_dataset("epfl-neuroai/multimodal-brain-scaling", "pretraining", split="results")
df = ds.to_pandas()

Other splits follow the same pattern (layer_search, finetuning, mapping).

To rehydrate the local CSV artifacts expected by the analysis code in the GitHub repository:

python -m mbs.download_artifacts

Citation

@inproceedings{gokce2026multimodal_brain_scaling,
  title     = {Multimodal Scaling Laws for Task \& Data-Optimized Models of Visual Cortex},
  author    = {Abdulkadir Gokce and Yingtian Tang and Martin Schrimpf},
  booktitle = {Forty-third International Conference on Machine Learning},
  year      = {2026},
  url       = {https://openreview.net/forum?id=OQ6jQHJPTT}
}

License

Released under the MIT License. Underlying neural datasets retain their original licenses; please refer to the respective sources before redistribution.