Class-Balanced Softmax: A Bayes Theory-Based Method for Long-Tailed Recognition

2026-07-27 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

Class-Balanced Softmax (CBS) is a novel logit adjustment method designed to improve deep learning model performance on imbalanced datasets, particularly in long-tailed recognition tasks. Traditional softmax classifiers and even Balanced Softmax often suffer from a "preference issue," where models exhibit significantly higher training error and larger generalization gaps for tail classes. Rooted in a theoretical Bayesian framework and a power-law assumption, CBS introduces a hyperparameter β ≥ 0 that modifies raw logits z_nc to z_nc+βlog|N_c|, effectively generalizing Balanced Softmax. This computationally inexpensive approach seamlessly integrates into existing pipelines. Extensive experiments on large-scale benchmarks like ImageNet-LT, iNaturalist2018, and an LVIS-based dataset demonstrate CBS's high scalability and superior performance, consistently achieving the highest average testing recall and effectively mitigating the preference issue by balancing gradients during training.

Key takeaway

For Machine Learning Engineers addressing long-tailed recognition, implementing Class-Balanced Softmax (CBS) provides a robust solution to the "preference issue" in imbalanced datasets. You should integrate this computationally inexpensive logit adjustment, z_nc+βlog|N_c|, into your deep learning pipelines. This will significantly improve average testing recall and balance performance across head and tail classes, outperforming existing methods like Balanced Softmax, particularly in large-scale, highly imbalanced scenarios.

Key insights

Class-Balanced Softmax (CBS) mitigates model preference for head classes by dynamically rebalancing gradients via a parameterized logit adjustment.

Principles

• Imbalanced data creates a "preference issue" in models.
• Logit adjustment rebalances class-wise gradient signals.
• Power-law functions model class frequency-prediction probability.

Method

CBS modifies raw logits z_nc to z_nc+βlog|N_c| using a calibrated power-law hyperparameter β ≥ 0 to balance gradients during back-propagation.

In practice

• Integrate CBS as a logit adjustment into existing pipelines.
• Calibrate β via grid search, starting in range [1.0, 1.3].
• Combine CBS with contrastive learning for enhanced performance.

Topics

• Long-tailed Recognition
• Imbalanced Classification
• Class-Balanced Softmax
• Logit Adjustment
• Gradient Balancing
• ImageNet-LT

Code references

• YihangZhu/Class-Balanced-Softmax

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Class-Balanced Softmax: A Bayes Theory-Based Method for Long-Tailed Recognition — Class-Balanced Softmax (CBS) improves long-tailed recognition by a Bayes-based logit adjustment, mitigating the "preference issue" in imbalanced datasets.
• Class-Specific Branch Attention for Mitigating Gradient Interference under Class Imbalance — Mitigating inter-class gradient interference in shared representations significantly improves minority-class learning under severe imbalance.
• SemEval-2026 Task 13: Fine-tuned CodeBERT with Stratified Balancing, Dynamic Threshold Optimization, and Logit Bias Correction for Robust Multi-Language AI Code Detection — Robust AI code detection requires addressing class imbalance and model overconfidence through specialized calibration techniques.
• Correcting Class Imbalance in Prior-Data Fitted Networks for Tabular Classification — PFNs' class imbalance issues are best addressed by simple thresholding or downsampling due to their unique calibration.
• Why Not Hyperparameter-Friendly Optimisation? A Monotonic Adaptive Norm Rescaling Approach For Long-Tailed Recognition — SAMN improves long-tailed recognition by enforcing monotonic per-class weight norms without sensitive hyperparameters.
• Model-Agnostic Meta Learning for Class Imbalance Adaptation — HAMR dynamically addresses class imbalance and data difficulty in NLP via adaptive weighting and neighborhood-aware resampling.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.