Perfectly Balanced: Improving Transfer and Robustness of Supervised Contrastive Learning

Authors: Mayee Chen, Daniel Y Fu, Avanika Narayan, Michael Zhang, Zhao Song, Kayvon Fatahalian, Christopher Re

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate THANOS on two tasks designed to evaluate how well it preserves subclasses: Coarse-to-fine transfer learning trains a model to classify superclasses but use the representations to distinguish subclasses. THANOS outperforms Sup Con by 11.1 points on average across 5 standard datasets. Worst-group robustness evaluates how well a model can identify underperforming sub-groups and maintain high performance on them. THANOS identifies underperforming sub-groups 7.7 points better than previous work (Sohoni et al., 2020) and achieves 4.7 points of lift on worst-group robustness across 3 datasets, setting state-of-the-art on Celeb A by 11.5 points.
Researcher Affiliation Collaboration 1Department of Computer Science, Stanford University 2Adobe Research.
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https://github.com/HazyResearch/thanos-code/.
Open Datasets Yes We use coarse versions of CIFAR10, CIFAR100, MNIST, and Tiny Image Net to study coarse-to-fine transfer. We use Waterbirds, ISIC and Celeb A for robustness (Sagawa et al., 2019; Codella et al., 2019; Liu et al., 2015; Sohoni et al., 2020).
Dataset Splits No The paper describes the datasets and training process but does not explicitly state the train/validation/test dataset splits used for reproduction, nor does it refer to specific predefined splits with proper citation for that split.
Hardware Specification Yes All transfer experiments were run using Tesla V100 machines.
Software Dependencies No We use the implementation in Py Torch Lightning Bolts2 (Falcon & Cho, 2020).
Experiment Setup Yes For the coarse dataset training, all models were trained for 600 epochs with an initial learning rate of 0.0003, a cosine annealing learning rate scheduler with Tmax set to 100 and the Adam W optimizer. A dropout rate of 0.05 was used. We did not use weight decay. All experiments were run using a batch size of 128 for both training and evaluation.