This repository hosts materials for our NeurIPS 2019 publication:
Kubilius*, Schrimpf*, et al. Brain-Like Object Recognition with High-Performing Shallow Recurrent ANNs. NeurIPS 2019 (oral)
* Equal contribution
This paper brings forward two major contributions:
- Brain-Score, a framework for evaluating models on integrative brain measurements. Brain-Score allows to quantify how similar models are to brain responses (neural and behavioral). The current Brain-Score leaderboard is available at Brain-Score.org. If you want to score your own model, use the Brain-Score repo
- CORnet-S, a shallow recurrent artificial neural network that is the current best model on Brain-Score. A PyTorch version of ImageNet pre-trained is available at CORnet repo.
Please cite this work as follows:
@inproceedings{KubiliusSchrimpf2019neurips,
title={Brain-Like Object Recognition with High-Performing Shallow Recurrent ANNs},
author={Kubilius, Jonas and Schrimpf, Martin and Kar, Kohitij and Hong, Ha and Majaj, Najib J and Rajalingham, Rishi and Issa, Elias B and Bashivan, Pouya and Prescott-Roy, Jonathan and Schmidt, Kailyn and Nayebi, Aran and Bear, Daniel and Yamins, Daniel L K and DiCarlo, James J},
booktitle={Advances in Neural Information Processing Systems},
year={2019}
}
We provide aggregated data sources for reproducing most of the figures in the paper. Run python figures.py gen_all
in order to generate all figures except Fig. 4 and 5. Data for Fig. 4 involves a comparison of many models; we chose to not package all that data. For Fig. 5, run python fig4.py prediction_vs_target
, but note that it will be recomputed from scratch and will therefore require multiple dependencies and may take a long time.
The data used in these figures has been computed using older versions of Brain-Score and thus may not perfectly reproduce when using the latest releases. We highly recommend using the latest release of Brain-Score (and the current scores in the leaderboard at Brain-Score.org) if you intend to report on your own data or models.