Research map / Label-efficient and robust learning / Recognition and long-tail

Separating common from salient patterns with Contrastive Representation Learning

ICLR 2024 · Label-efficient and robust learning › Recognition and long-tail · approach: Contrastive representation theory · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SepCLR, an InfoMax-based contrastive framework that separates common from target-specific (salient) factors between background and target datasets, using alignment/uniformity losses and a kernel-based joint entropy maximization to enforce common-salient independence.

Problem

Contrastive Analysis: separating factors shared between a background and target dataset from salient factors present only in the target, where existing VAE-based methods learn poor semantic representations.

Approach

Reformulates contrastive analysis under the InfoMax principle, estimates the common and salient MI terms with contrastive alignment and uniformity losses (background-contrasting InfoNCE), and enforces common-salient independence by maximizing joint entropy H(c,s) via kernel density estimation rather than minimizing mutual information.

Key results

• On Digits on CIFAR-10, SepCLR-k-JEM reaches delta_tot 32.0 versus 81.2 for the best VAE baseline (SepVAE).
• On CelebA accessories, SepCLR-k-JEM achieves delta_tot 41.16 versus 64.60 for the best baseline (cVAE variant at 64.60).
• On schizophrenia, SepCLR-k-JEM gives the highest salient diagnosis accuracy of 79.90 versus 79.15 for SepVAE, and best site and sex separation.

Measured effects reported in the paper

Surprising findings

• Adding a reconstruction loss during training degrades performance, according to the authors.
• Minimizing I(c;s) via variational MI estimators or minimizing H(c)+H(s) underperforms enforcing I(c;s)=0 through joint entropy maximization.

Limitations

• Identifiability of the true generative factors is not guaranteed; SepCLR separates but does not ensure all factors are recovered.
• Performance depends on hyperparameter weights (lambda_C, lambda_S, lambda) that vary across datasets.
• Reconstruction loss degrades performance, so no generator is included.

Experimental setup

Datasets: CIFAR-10 MNIST CelebA dSprites CheXpert ODIR schizophrenia MRI (VBM)

Benchmarks: Digits on CIFAR-10 CelebA accessories schizophrenia neuroimaging CheXpert ODIR dSprites disentanglement

Metrics: balanced accuracy AUC MAE R^2 delta_tot

Baselines: cVAE ConVAE MM-cVAE SepVAE vCLUB vL1out vUB TC

Builds on: InfoNCE SupCon SimCLR contrastive VAE DeepSAD

Closest papers by meaning

• Variational Supervised Contrastive Learning (NeurIPS 2025 · similarity 0.88)
• Rethinking Minimal Sufficient Representation in Contrastive Learning (CVPR 2022 · similarity 0.88)
• RényiCL: Contrastive Representation Learning with Skew Rényi Divergence (NeurIPS 2022 · similarity 0.87)
• Multi-modal contrastive learning adapts to intrinsic dimensions of shared latent variables (NeurIPS 2025 · similarity 0.87)
• A Statistical Theory of Contrastive Learning via Approximate Sufficient Statistics (NeurIPS 2025 · similarity 0.86)
• An Investigation of Representation and Allocation Harms in Contrastive Learning (ICLR 2024 · similarity 0.86)