Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Understanding Invariance via Feedforward Inversion of Discriminatively Trained Classifiers
Authors: Piotr Teterwak, Chiyuan Zhang, Dilip Krishnan, Michael C Mozer
ICML 2021 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We train the generator on TPUv3 accelerators using the same Image Net (Russakovsky et al., 2015) training set that the classifiers are trained on, and evaluate using the test set. |
| Researcher Affiliation | Collaboration | 1Presently at Boston University; work was begun while author was an AI Resident at Google Research 2Google Research 3University of Colorado, Boulder. |
| Pseudocode | No | No structured pseudocode or algorithm blocks were found in the paper. |
| Open Source Code | No | We provide pre-trained models and visualizations at https://sites.google.com/view/ understanding-invariance/home. |
| Open Datasets | Yes | We train the generator on TPUv3 accelerators using the same Image Net (Russakovsky et al., 2015) training set that the classifiers are trained on, and evaluate using the test set. |
| Dataset Splits | No | No specific dataset split information (percentages, sample counts, or explicit splitting methodology) for training, validation, and test sets was found for general reproducibility. While 'validation sample logit vectors' are mentioned for a specific analysis, a complete split breakdown is not provided. |
| Hardware Specification | Yes | We train the generator on TPUv3 accelerators using the same Image Net (Russakovsky et al., 2015) training set that the classifiers are trained on, and evaluate using the test set. |
| Software Dependencies | No | No specific version numbers for key software components (e.g., libraries, frameworks) are provided, only names like 'Big GAN implementation' and 'Caffe'. |
| Experiment Setup | No | While some experiment-specific parameters (e.g., FGSM attack strength ǫ = 0.1, noise for logit perturbation N(µ = 0, σ2 = 0.55)) are mentioned, the main training configurations and hyperparameters (learning rate, batch size, epochs, optimizer) are stated to be 'as described in the Supplementary Materials', not explicitly in the main text. |