ConR: Contrastive Regularizer for Deep Imbalanced Regression

TL;DR

ConR introduces a continuous contrastive regularizer for deep imbalanced regression that transfers inter-label similarities into the feature space to prevent minority samples from collapsing into majority regions. It employs dynamic anchor selection, targeted negative sampling, and relative pushing weights tied to label density, optimizing a loss $L_{sum} = \alpha L_{R} + \beta L_{ConR}$ with $L_{ConR}$ defined in a continuous adaptation of infoNCE. Theoretical analysis provides an upper bound linking minimizing $L_{ConR}$ to reduced probability of incorrect minority labeling, while extensive experiments on AgeDB-DIR, IMDB-WIKI-DIR, NYUD2-DIR, and MPIIGaze-DIR show consistent improvements over state-of-the-art baselines. ConR is orthogonal to existing DIR approaches and scales to high-dimensional label spaces, delivering strong gains with modest computational overhead. The work offers both practical applicability and a conceptual shift toward biased-robust contrastive learning for continuous regression tasks, with code available at the provided repository.

Abstract

Imbalanced distributions are ubiquitous in real-world data. They create constraints on Deep Neural Networks to represent the minority labels and avoid bias towards majority labels. The extensive body of imbalanced approaches address categorical label spaces but fail to effectively extend to regression problems where the label space is continuous. Local and global correlations among continuous labels provide valuable insights towards effectively modelling relationships in feature space. In this work, we propose ConR, a contrastive regularizer that models global and local label similarities in feature space and prevents the features of minority samples from being collapsed into their majority neighbours. ConR discerns the disagreements between the label space and feature space and imposes a penalty on these disagreements. ConR addresses the continuous nature of label space with two main strategies in a contrastive manner: incorrect proximities are penalized proportionate to the label similarities and the correct ones are encouraged to model local similarities. ConR consolidates essential considerations into a generic, easy-to-integrate, and efficient method that effectively addresses deep imbalanced regression. Moreover, ConR is orthogonal to existing approaches and smoothly extends to uni- and multi-dimensional label spaces. Our comprehensive experiments show that ConR significantly boosts the performance of all the state-of-the-art methods on four large-scale deep imbalanced regression benchmarks. Our code is publicly available in https://github.com/BorealisAI/ConR.

Figures (15)

• Figure 1: Key insights of ConR. a) Without ConR, it is common to have minority examples mixed with majority examples. b) ConR selects the sample with confusion around it as an anchor and adjusts the feature space with relative contrastive learning. c) Reduced prediction error.
• Figure 2: The framework of ConR is to translate label similarities to the feature space. a) Per each augmented sample, ConR selects positive and negative pairs with regard to the label similarities and prediction similarities. b) ConR pulls positive pairs together while pushing away negative pairs regarding their label similarities and label distribution for the anchor. In this way, the minority anchor pushes negative samples harder. The pushing weight is inversely relative to the label similarities.
• Figure 3: comparison on RMSE results by adding ConR on top of the baselines for NYUD2-DIR benchmark.
• Figure 4: Feature visualization on AgeDB-DIR for (a) VANILLA, (b) FDS, (c) LDS and (d) ConR.
• Figure 5: Ablation study on the similarity threshold $\omega$. (a) and (b) compares the learnt feature space for similarity threshold of $2$ and $1$ respectively. (c) Comparison of different choices of $\omega$ in terms of MAE.