{"id":15468996,"url":"https://github.com/ajaysub110/satbench","last_synced_at":"2025-06-20T03:05:48.053Z","repository":{"id":37794702,"uuid":"501396766","full_name":"ajaysub110/satbench","owner":"ajaysub110","description":"Benchmarking the speed-accuracy tradeoff in object recognition by humans and dynamic neural networks","archived":false,"fork":false,"pushed_at":"2022-10-17T22:36:23.000Z","size":24654,"stargazers_count":0,"open_issues_count":0,"forks_count":1,"subscribers_count":1,"default_branch":"main","last_synced_at":"2025-06-03T22:55:06.544Z","etag":null,"topics":[],"latest_commit_sha":null,"homepage":"","language":"Jupyter Notebook","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":null,"status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/ajaysub110.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":null,"code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null}},"created_at":"2022-06-08T20:12:54.000Z","updated_at":"2022-07-19T23:36:29.000Z","dependencies_parsed_at":"2022-06-23T17:58:11.861Z","dependency_job_id":null,"html_url":"https://github.com/ajaysub110/satbench","commit_stats":null,"previous_names":[],"tags_count":0,"template":false,"template_full_name":null,"purl":"pkg:github/ajaysub110/satbench","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/ajaysub110%2Fsatbench","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/ajaysub110%2Fsatbench/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/ajaysub110%2Fsatbench/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/ajaysub110%2Fsatbench/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/ajaysub110","download_url":"https://codeload.github.com/ajaysub110/satbench/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/ajaysub110%2Fsatbench/sbom","host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":260868849,"owners_count":23074959,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2022-07-04T15:15:14.044Z","host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":[],"created_at":"2024-10-02T01:48:43.268Z","updated_at":"2025-06-20T03:05:43.039Z","avatar_url":"https://github.com/ajaysub110.png","language":"Jupyter Notebook","funding_links":[],"categories":[],"sub_categories":[],"readme":"# SATBench: Benchmarking the speed-accuracy tradeoff in object recognition by humans and dynamic neural networks\n\n**This repository contains code used in the paper:**\n\n[SATBench: Benchmarking the speed-accuracy tradeoff in object recognition by humans and dynamic neural networks](https://arxiv.org/abs/2206.08427) by Ajay Subramanian, Omkar Kumbhar, Elena Sizikova, Najib J. Majaj, Denis G. Pelli (New York University, 2022).\n\nOur contributions are as follows:\n* We present large-scale (148 human observers), public dataset on timed ImageNet [1] object recognition with 16 categories, across color, grayscale, 3 noise and 3 blur conditions. For each condition, we tested human performance for 5 reaction time (RT) values. This data provides a benchmark for the human speed-accuracy tradeoff and is specifically intended to facilitate comparison between neural networks and humans on timed object recognition.\n* We present comparable benchmarks for dynamic neural networks, a class of networks capable of inference-time adaptive computation.\n* We perform an extensive quantitative comparison between speed-accuracy tradeoffs in humans and four dynamic neural networks. To do so, we propose three novel metrics: RMSE between SAT curves, category-wise correlation, and steepness which ease model-human comparison.\n\n![](assets/human-network-sat.jpeg)\n\n## Table of Contents\n1. Dataset\n2. Code\n3. Citation\n4. References\n\n## Dataset\nOur human dataset is collected using a reaction time paradigm proposed by McElree \u0026 Carrasco [2] where observers are forced to respond at a beep which sounds at a specific time after target presentation. Varying the beep interval across several blocks helps us collect object recognition data across different reaction times (`500ms`, `900ms`, `1100ms`, `1300ms`, `1500ms`). We evaluate dynamic neural networks using the same paradigm with computational FLOPs used as an analog for reaction time.\n\nHuman dataset and network results can be found at https://osf.io/2cpmb/. Download and unzip `human-data.zip` and `model_data.zip` for human and network data respectively.\n\n## Code\n1. Code to generate image dataset used in the paper is available in the `generate_images` directory.\n2. JSON files corresponding to LabJS studies used to collect human data are available in the `human_data_collection` directory.\n3. Code used to analyze human data is available as notebooks in the `human_data_analysis` directory.\n4. We benchmark 4 dynamic neural network models - MSDNet [2] , SCAN [3], Cascaded-Nets (CNets) [4]  and ConvRNN [3] on our dataset. The following table mentions the scripts to be used for training and inference of each model. We used code for each model from existing/official implementations (links given below):\n  - MSDNet: https://github.com/kalviny/MSDNet-PyTorch\n  - SCAN: https://github.com/ArchipLab-LinfengZhang/pytorch-scalable-neural-networks\n  - ConvRNN: https://github.com/cjspoerer/rcnn-sat\n  - CNet: https://github.com/michael-iuzzolino/CascadedNets\n\n  Code for each model is available in a subdirectory with the model's name. Links to pretrained networks will be added to respective README.md files upon publication.\n\n## Citation\n```\n@article{subramanian2022satbench,\n  title={SATBench: Benchmarking the speed-accuracy tradeoff in object recognition by humans and dynamic neural networks},\n  author={Subramanian, Ajay and Price, Sara and Kumbhar, Omkar and Sizikova, Elena and Majaj, Najib J and Pelli, Denis G},\n  journal={arXiv preprint arXiv:2206.08427},\n  year={2022}\n}\n```\n\n## References\n[1] Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., ... \u0026 Fei-Fei, L. (2015). Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3), 211-252.\n\n[2] Huang, G., Chen, D., Li, T., Wu, F., Van Der Maaten, L., \u0026 Weinberger, K. Q. (2017). Multi-scale dense networks for resource efficient image classification. arXiv preprint arXiv:1703.09844.\n\n[3] Zhang, L., Tan, Z., Song, J., Chen, J., Bao, C., \u0026 Ma, K. (2019). Scan: A scalable neural networks framework towards compact and efficient models. Advances in Neural Information Processing Systems, 32.\n\n[4] Iuzzolino, M., Mozer, M. C., \u0026 Bengio, S. (2021). Improving Anytime Prediction with Parallel Cascaded Networks and a Temporal-Difference Loss. Advances in Neural Information Processing Systems, 34.\n\n[5] Spoerer, C. J., Kietzmann, T. C., Mehrer, J., Charest, I., \u0026 Kriegeskorte, N. (2020). Recurrent neural networks can explain flexible trading of speed and accuracy in biological vision. PLoS computational biology, 16(10), e1008215.\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fajaysub110%2Fsatbench","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fajaysub110%2Fsatbench","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fajaysub110%2Fsatbench/lists"}