Reproduce “ConR: Contrastive Regularizer for Deep Imbalanced Regression” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Mahsa Keramati, Lili Meng, R. David Evans

CONR (Contrastive Regularizer) adapts contrastive learning to deep imbalanced regression by explicitly modeling local and global label relationships in continuous label spaces. It boosts state-of-the-art baselines on four datasets (age, depth, gaze estimation), particularly for minority samples, and learns more effective feature representations.

Key insights

  • Feature collapse, where minority samples become indistinguishable from majority ones, is a primary cause of poor performance in imbalanced regression, and explicitly targeting this via contrastive learning is effective.
  • Dynamically identifying and penalizing 'false positives' (dissimilar labels with similar predictions) using a contrastive loss, especially with amplified penalties for minority samples, improves generalization.
  • The relative pushing mechanism, which assigns pushing power based on inverse label frequency and label dissimilarity, is crucial for effectively rebalancing the feature space and preventing collapse.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.