Scores on benchmarks
Model rank shown below is with respect to all public models.| .190 |
average_vision
rank 225
111 benchmarks |
|
| .020 |
neural_vision
rank 496
68 benchmarks |
|
| .009 |
V1
rank 503
31 benchmarks |
|
| .063 |
Coggan2024_fMRI.V1-rdm
v1
rank 103
|
|
|
24 images 
|
||
| .030 |
V2
rank 480
8 benchmarks |
|
| .148 |
Coggan2024_fMRI.V2-rdm
v1
rank 70
|
|
|
24 images 
|
||
| .006 |
V4
rank 498
13 benchmarks |
|
| .060 |
Coggan2024_fMRI.V4-rdm
v1
rank 82
|
|
|
24 images 
|
||
| .037 |
IT
rank 458
16 benchmarks |
|
| .483 |
Coggan2024_fMRI.IT-rdm
v1
rank 84
|
|
|
24 images 
|
||
| .359 |
behavior_vision
rank 108
43 benchmarks |
|
| .542 |
Rajalingham2018-i2n
v2
[reference]
rank 86
|
|
|
match-to-sample task
240 images 
|
||
| .319 |
Geirhos2021-error_consistency
[reference]
rank 93
17 benchmarks |
|
| .624 |
Geirhos2021colour-error_consistency
v1
[reference]
rank 52
|
|
|
640 images 
|
||
| .246 |
Geirhos2021contrast-error_consistency
v1
[reference]
rank 111
|
|
|
800 images 
|
||
| .255 |
Geirhos2021cueconflict-error_consistency
v1
[reference]
rank 97
|
|
|
1280 images 
|
||
| .060 |
Geirhos2021edge-error_consistency
v1
[reference]
rank 247
|
|
|
160 images 
|
||
| .470 |
Geirhos2021eidolonI-error_consistency
v1
[reference]
rank 91
|
|
|
800 images 
|
||
| .524 |
Geirhos2021eidolonII-error_consistency
v1
[reference]
rank 63
|
|
|
640 images 
|
||
| .433 |
Geirhos2021eidolonIII-error_consistency
v1
[reference]
rank 81
|
|
|
480 images 
|
||
| .546 |
Geirhos2021falsecolour-error_consistency
v1
[reference]
rank 57
|
|
|
560 images 
|
||
| .031 |
Geirhos2021highpass-error_consistency
v1
[reference]
rank 249
|
|
|
640 images 
|
||
| .270 |
Geirhos2021lowpass-error_consistency
v1
[reference]
rank 93
|
|
|
800 images 
|
||
| .210 |
Geirhos2021phasescrambling-error_consistency
v1
[reference]
rank 95
|
|
|
640 images 
|
||
| .260 |
Geirhos2021powerequalisation-error_consistency
v1
[reference]
rank 80
|
|
|
560 images 
|
||
| .236 |
Geirhos2021rotation-error_consistency
v1
[reference]
rank 79
|
|
|
960 images 
|
||
| .445 |
Geirhos2021silhouette-error_consistency
v1
[reference]
rank 154
|
|
|
160 images 
|
||
| .131 |
Geirhos2021sketch-error_consistency
v1
[reference]
rank 110
|
|
|
800 images 
|
||
| .456 |
Geirhos2021stylized-error_consistency
v1
[reference]
rank 69
|
|
|
800 images 
|
||
| .231 |
Geirhos2021uniformnoise-error_consistency
v1
[reference]
rank 118
|
|
|
800 images 
|
||
| .157 |
Baker2022
rank 164
3 benchmarks |
|
| .470 |
Baker2022fragmented-accuracy_delta
v1
[reference]
rank 114
|
|
|
716 images 
|
||
| .000 |
Baker2022frankenstein-accuracy_delta
v1
[reference]
rank 175
|
|
|
716 images 
|
||
| .000 |
Baker2022inverted-accuracy_delta
v1
[reference]
rank 71
|
|
|
360 images 
|
||
| .162 |
BMD2024
rank 133
4 benchmarks |
|
| .166 |
BMD2024.dotted_1Behavioral-accuracy_distance
v1
rank 101
|
|
|
100 images 
|
||
| .126 |
BMD2024.dotted_2Behavioral-accuracy_distance
v1
rank 132
|
|
|
100 images 
|
||
| .186 |
BMD2024.texture_1Behavioral-accuracy_distance
v1
rank 116
|
|
|
100 images 
|
||
| .168 |
BMD2024.texture_2Behavioral-accuracy_distance
v1
rank 134
|
|
|
100 images 
|
||
| .444 |
Ferguson2024
[reference]
rank 163
14 benchmarks |
|
| .353 |
Ferguson2024circle_line-value_delta
v1
[reference]
rank 125
|
|
|
2_way_afc task
48 images 
|
||
| .719 |
Ferguson2024color-value_delta
v1
[reference]
rank 143
|
|
|
2_way_afc task
48 images 
|
||
| .177 |
Ferguson2024convergence-value_delta
v1
[reference]
rank 224
|
|
|
2_way_afc task
48 images 
|
||
| .060 |
Ferguson2024eighth-value_delta
v1
[reference]
rank 216
|
|
|
2_way_afc task
48 images 
|
||
| .725 |
Ferguson2024gray_easy-value_delta
v1
[reference]
rank 72
|
|
|
2_way_afc task
48 images 
|
||
| .138 |
Ferguson2024gray_hard-value_delta
v1
[reference]
rank 265
|
|
|
2_way_afc task
48 images 
|
||
| .571 |
Ferguson2024half-value_delta
v1
[reference]
rank 126
|
|
|
2_way_afc task
48 images 
|
||
| .288 |
Ferguson2024juncture-value_delta
v1
[reference]
rank 96
|
|
|
2_way_afc task
48 images 
|
||
| .161 |
Ferguson2024lle-value_delta
v1
[reference]
rank 236
|
|
|
2_way_afc task
48 images 
|
||
| .583 |
Ferguson2024llh-value_delta
v1
[reference]
rank 121
|
|
|
2_way_afc task
48 images 
|
||
| .950 |
Ferguson2024quarter-value_delta
v1
[reference]
rank 16
|
|
|
2_way_afc task
48 images 
|
||
| .102 |
Ferguson2024round_f-value_delta
v1
[reference]
rank 248
|
|
|
2_way_afc task
48 images 
|
||
| .409 |
Ferguson2024round_v-value_delta
v1
[reference]
rank 164
|
|
|
2_way_afc task
48 images 
|
||
| .975 |
Ferguson2024tilted_line-value_delta
v1
[reference]
rank 29
|
|
|
2_way_afc task
48 images 
|
||
| .307 |
Hebart2023-match
v1
rank 122
|
|
|
1854 images 
|
||
| .573 |
Maniquet2024
rank 125
2 benchmarks |
|
| .500 |
Maniquet2024-confusion_similarity
v1
[reference]
rank 138
|
|
|
13600 images 
|
||
| .646 |
Maniquet2024-tasks_consistency
v1
[reference]
rank 148
|
|
|
13600 images 
|
||
| .371 |
Coggan2024_behavior-ConditionWiseAccuracySimilarity
v1
rank 97
|
|
|
22560 images 
|
||
| .396 |
engineering_vision
rank 99
25 benchmarks |
|
| .749 |
ImageNet-top1
v1
[reference]
rank 70
|
|
|
50000 images 
|
||
| .445 |
ImageNet-C-top1
[reference]
rank 62
4 benchmarks |
|
| .360 |
ImageNet-C-noise-top1
v2
[reference]
rank 104
|
|
|
||
| .412 |
ImageNet-C-blur-top1
v2
[reference]
rank 44
|
|
|
||
| .499 |
ImageNet-C-weather-top1
v2
[reference]
rank 65
|
|
|
||
| .507 |
ImageNet-C-digital-top1
v2
[reference]
rank 77
|
|
|
||
| .584 |
Geirhos2021-top1
[reference]
rank 107
17 benchmarks |
|
| .980 |
Geirhos2021colour-top1
v1
[reference]
rank 83
|
|
|
640 images 
|
||
| .829 |
Geirhos2021contrast-top1
v1
[reference]
rank 131
|
|
|
800 images 
|
||
| .207 |
Geirhos2021cueconflict-top1
v1
[reference]
rank 159
|
|
|
1280 images 
|
||
| .206 |
Geirhos2021edge-top1
v1
[reference]
rank 224
|
|
|
160 images 
|
||
| .526 |
Geirhos2021eidolonI-top1
v1
[reference]
rank 100
|
|
|
800 images 
|
||
| .570 |
Geirhos2021eidolonII-top1
v1
[reference]
rank 70
|
|
|
640 images 
|
||
| .550 |
Geirhos2021eidolonIII-top1
v1
[reference]
rank 124
|
|
|
480 images 
|
||
| .954 |
Geirhos2021falsecolour-top1
v1
[reference]
rank 110
|
|
|
560 images 
|
||
| .294 |
Geirhos2021highpass-top1
v1
[reference]
rank 218
|
|
|
640 images 
|
||
| .478 |
Geirhos2021lowpass-top1
v1
[reference]
rank 82
|
|
|
800 images 
|
||
| .656 |
Geirhos2021phasescrambling-top1
v1
[reference]
rank 85
|
|
|
640 images 
|
||
| .793 |
Geirhos2021powerequalisation-top1
v1
[reference]
rank 93
|
|
|
560 images 
|
||
| .729 |
Geirhos2021rotation-top1
v1
[reference]
rank 89
|
|
|
960 images 
|
||
| .563 |
Geirhos2021silhouette-top1
v1
[reference]
rank 71
|
|
|
160 images 
|
||
| .668 |
Geirhos2021sketch-top1
v1
[reference]
rank 80
|
|
|
800 images 
|
||
| .416 |
Geirhos2021stylized-top1
v1
[reference]
rank 110
|
|
|
800 images 
|
||
| .516 |
Geirhos2021uniformnoise-top1
v1
[reference]
rank 103
|
|
|
800 images 
|
||
| .204 |
Hermann2020
[reference]
rank 200
2 benchmarks |
|
| .245 |
Hermann2020cueconflict-shape_bias
v1
[reference]
rank 202
|
|
|
||
| .163 |
Hermann2020cueconflict-shape_match
v1
[reference]
rank 167
|
|
|
||
How to use
from brainscore_vision import load_model
model = load_model("resnet_152_v1")
model.start_task(...)
model.start_recording(...)
model.look_at(...)
Brain Encoding Response Generator (BERG)
Through the BERG you can easily generate neural responses to images of your choice using any Brain-Score vision model.
For more information on how to use BERG, see the documentation and tutorial.
Benchmarks bibtex
@inproceedings{santurkar2019computer,
title={Computer Vision with a Single (Robust) Classifier},
author={Shibani Santurkar and Dimitris Tsipras and Brandon Tran and Andrew Ilyas and Logan Engstrom and Aleksander Madry},
booktitle={ArXiv preprint arXiv:1906.09453},
year={2019}
}
@article {Rajalingham240614,
author = {Rajalingham, Rishi and Issa, Elias B. and Bashivan, Pouya and Kar, Kohitij and Schmidt, Kailyn and DiCarlo, James J.},
title = {Large-scale, high-resolution comparison of the core visual object recognition behavior of humans, monkeys, and state-of-the-art deep artificial neural networks},
elocation-id = {240614},
year = {2018},
doi = {10.1101/240614},
publisher = {Cold Spring Harbor Laboratory},
abstract = {Primates{ extemdash}including humans{ extemdash}can typically recognize objects in visual images at a glance even in the face of naturally occurring identity-preserving image transformations (e.g. changes in viewpoint). A primary neuroscience goal is to uncover neuron-level mechanistic models that quantitatively explain this behavior by predicting primate performance for each and every image. Here, we applied this stringent behavioral prediction test to the leading mechanistic models of primate vision (specifically, deep, convolutional, artificial neural networks; ANNs) by directly comparing their behavioral signatures against those of humans and rhesus macaque monkeys. Using high-throughput data collection systems for human and monkey psychophysics, we collected over one million behavioral trials for 2400 images over 276 binary object discrimination tasks. Consistent with previous work, we observed that state-of-the-art deep, feed-forward convolutional ANNs trained for visual categorization (termed DCNNIC models) accurately predicted primate patterns of object-level confusion. However, when we examined behavioral performance for individual images within each object discrimination task, we found that all tested DCNNIC models were significantly non-predictive of primate performance, and that this prediction failure was not accounted for by simple image attributes, nor rescued by simple model modifications. These results show that current DCNNIC models cannot account for the image-level behavioral patterns of primates, and that new ANN models are needed to more precisely capture the neural mechanisms underlying primate object vision. To this end, large-scale, high-resolution primate behavioral benchmarks{ extemdash}such as those obtained here{ extemdash}could serve as direct guides for discovering such models.SIGNIFICANCE STATEMENT Recently, specific feed-forward deep convolutional artificial neural networks (ANNs) models have dramatically advanced our quantitative understanding of the neural mechanisms underlying primate core object recognition. In this work, we tested the limits of those ANNs by systematically comparing the behavioral responses of these models with the behavioral responses of humans and monkeys, at the resolution of individual images. Using these high-resolution metrics, we found that all tested ANN models significantly diverged from primate behavior. Going forward, these high-resolution, large-scale primate behavioral benchmarks could serve as direct guides for discovering better ANN models of the primate visual system.},
URL = {https://www.biorxiv.org/content/early/2018/02/12/240614},
eprint = {https://www.biorxiv.org/content/early/2018/02/12/240614.full.pdf},
journal = {bioRxiv}
}
@article{geirhos2021partial,
title={Partial success in closing the gap between human and machine vision},
author={Geirhos, Robert and Narayanappa, Kantharaju and Mitzkus, Benjamin and Thieringer, Tizian and Bethge, Matthias and Wichmann, Felix A and Brendel, Wieland},
journal={Advances in Neural Information Processing Systems},
volume={34},
year={2021},
url={https://openreview.net/forum?id=QkljT4mrfs}
}
@article{BAKER2022104913,
title = {Deep learning models fail to capture the configural nature of human shape perception},
journal = {iScience},
volume = {25},
number = {9},
pages = {104913},
year = {2022},
issn = {2589-0042},
doi = {https://doi.org/10.1016/j.isci.2022.104913},
url = {https://www.sciencedirect.com/science/article/pii/S2589004222011853},
author = {Nicholas Baker and James H. Elder},
keywords = {Biological sciences, Neuroscience, Sensory neuroscience},
abstract = {Summary
A hallmark of human object perception is sensitivity to the holistic configuration of the local shape features of an object. Deep convolutional neural networks (DCNNs) are currently the dominant models for object recognition processing in the visual cortex, but do they capture this configural sensitivity? To answer this question, we employed a dataset of animal silhouettes and created a variant of this dataset that disrupts the configuration of each object while preserving local features. While human performance was impacted by this manipulation, DCNN performance was not, indicating insensitivity to object configuration. Modifications to training and architecture to make networks more brain-like did not lead to configural processing, and none of the networks were able to accurately predict trial-by-trial human object judgements. We speculate that to match human configural sensitivity, networks must be trained to solve a broader range of object tasks beyond category recognition.}
}
@misc{ferguson_ngo_lee_dicarlo_schrimpf_2024,
title={How Well is Visual Search Asymmetry predicted by a Binary-Choice, Rapid, Accuracy-based Visual-search, Oddball-detection (BRAVO) task?},
url={osf.io/5ba3n},
DOI={10.17605/OSF.IO/5BA3N},
publisher={OSF},
author={Ferguson, Michael E, Jr and Ngo, Jerry and Lee, Michael and DiCarlo, James and Schrimpf, Martin},
year={2024},
month={Jun}
}
@article {Maniquet2024.04.02.587669,
author = {Maniquet, Tim and de Beeck, Hans Op and Costantino, Andrea Ivan},
title = {Recurrent issues with deep neural network models of visual recognition},
elocation-id = {2024.04.02.587669},
year = {2024},
doi = {10.1101/2024.04.02.587669},
publisher = {Cold Spring Harbor Laboratory},
URL = {https://www.biorxiv.org/content/early/2024/04/10/2024.04.02.587669},
eprint = {https://www.biorxiv.org/content/early/2024/04/10/2024.04.02.587669.full.pdf},
journal = {bioRxiv}
}
@INPROCEEDINGS{5206848,
author={J. {Deng} and W. {Dong} and R. {Socher} and L. {Li} and {Kai Li} and {Li Fei-Fei}},
booktitle={2009 IEEE Conference on Computer Vision and Pattern Recognition},
title={ImageNet: A large-scale hierarchical image database},
year={2009},
volume={},
number={},
pages={248-255},
}
@ARTICLE{Hendrycks2019-di,
title = "Benchmarking Neural Network Robustness to Common Corruptions
and Perturbations",
author = "Hendrycks, Dan and Dietterich, Thomas",
abstract = "In this paper we establish rigorous benchmarks for image
classifier robustness. Our first benchmark, ImageNet-C,
standardizes and expands the corruption robustness topic,
while showing which classifiers are preferable in
safety-critical applications. Then we propose a new dataset
called ImageNet-P which enables researchers to benchmark a
classifier's robustness to common perturbations. Unlike
recent robustness research, this benchmark evaluates
performance on common corruptions and perturbations not
worst-case adversarial perturbations. We find that there are
negligible changes in relative corruption robustness from
AlexNet classifiers to ResNet classifiers. Afterward we
discover ways to enhance corruption and perturbation
robustness. We even find that a bypassed adversarial defense
provides substantial common perturbation robustness.
Together our benchmarks may aid future work toward networks
that robustly generalize.",
month = mar,
year = 2019,
archivePrefix = "arXiv",
primaryClass = "cs.LG",
eprint = "1903.12261",
url = "https://arxiv.org/abs/1903.12261"
}
@article{hermann2020origins,
title={The origins and prevalence of texture bias in convolutional neural networks},
author={Hermann, Katherine and Chen, Ting and Kornblith, Simon},
journal={Advances in Neural Information Processing Systems},
volume={33},
pages={19000--19015},
year={2020},
url={https://proceedings.neurips.cc/paper/2020/hash/db5f9f42a7157abe65bb145000b5871a-Abstract.html}
}