Debiased Contrastive Learning

Authors: Ching-Yao Chuang, Joshua Robinson, Yen-Chen Lin, Antonio Torralba, Stefanie Jegelka

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, the proposed objective consistently outperforms the state-of-the-art for representation learning in vision, language, and reinforcement learning benchmarks.
Researcher Affiliation Academia Ching-Yao Chuang, Joshua Robinson, Lin Yen-Chen Antonio Torralba, Stefanie Jegelka CSAIL, Massachusetts Institute of Technology Cambridge, MA 02139, USA {cychuang, joshrob, yenchenl, torralba, stefje}@mit.edu
Pseudocode Yes Figure 3: Pseudocode for debiased objective with M = 1. The implementation only requires a small modification of the code. We can simply extend the code to debiased objective with M > 1 by changing the pos in line 8 with an average of exponentials for M positive samples.
Open Source Code Yes The code is available at https://github.com/chingyaoc/DCL.
Open Datasets Yes First, for CIFAR10 [27] and STL10 [7], we implement Sim CLR [2] with Res Net-50 [17] as the encoder architecture and use the Adam optimizer [23] with learning rate 0.001. ... Following [40], we test our approach on Image Net-100, a randomly chosen subset of 100 classes of Imagenet. ... We use the Book Corpus dataset [25]...
Dataset Splits Yes All the models are trained for 400 epochs and evaluated by training a linear classifier after fixing the learned embedding. ... 10-fold cross validation is used in testing the performance for binary classification tasks (MR, CR, SUBJ, MPQA).
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models, processor types, or memory amounts used for running its experiments.
Software Dependencies No The paper mentions software components such as 'Sim CLR', 'Res Net-50', and 'Adam optimizer', but does not provide specific version numbers for any of them.
Experiment Setup Yes Following [2], we set the temperature to t = 0.5 and the dimension of the latent vector to 128. All the models are trained for 400 epochs and evaluated by training a linear classifier after fixing the learned embedding. ... use the Adam optimizer [23] with learning rate 0.001.