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

2026-07-24 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

Class-Balanced Softmax (CBS) is a new method designed to improve deep learning model performance on imbalanced datasets, specifically for long-tailed recognition. It addresses the limitations of existing rebalancing techniques like Balanced Softmax, which often yield low accuracy for tail classes. Rooted in a theoretical Bayesian framework and a heuristic power-law assumption, CBS functions as a simple, computationally inexpensive logit adjustment that integrates easily into current pipelines. The method also characterizes and mitigates a "preference issue," a phenomenon where models trained on imbalanced data exhibit higher training error and a larger generalization gap for data-limited classes. Extensive experiments on large-scale benchmarks demonstrate CBS's scalability and superior performance compared to other methods, including Balanced Softmax.

Key takeaway

For Machine Learning Engineers developing models for real-world, imbalanced datasets, adopting Class-Balanced Softmax (CBS) is crucial. It directly addresses the "preference issue" and significantly improves accuracy for underrepresented "tail" classes, outperforming prior methods like Balanced Softmax. You should integrate this computationally inexpensive logit adjustment into your existing pipelines to enhance model generalization and fairness.

Key insights

Class-Balanced Softmax (CBS) improves long-tailed recognition by a Bayes-based logit adjustment, mitigating the "preference issue" in imbalanced datasets.

Principles

• Imbalanced datasets degrade traditional softmax performance.
• Balanced Softmax has limitations for tail classes.
• Models show a "preference issue" for limited data classes.

Method

CBS is a simple logit adjustment derived from a theoretical Bayesian framework and a heuristic power-law assumption, easily integrated into existing deep learning pipelines.

In practice

• Integrate CBS into existing deep learning pipelines.
• Apply CBS for long-tailed recognition tasks.
• Use CBS to improve tail class accuracy.

Topics

• Long-Tailed Recognition
• Imbalanced Datasets
• Class-Balanced Softmax
• Bayesian Framework
• Logit Adjustment
• Deep Learning Classification

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

Related on AIssential

• Class-Balanced Softmax: A Bayes Theory-Based Method for Long-Tailed Recognition — Class-Balanced Softmax (CBS) mitigates model preference for head classes by dynamically rebalancing gradients via a parameterized logit adjustment.
• Practical estimation of the optimal classification error with soft labels and calibration — Isotonic calibration enables consistent Bayes error estimation even with corrupted soft labels, provided their order is preserved.
• 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.
• 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.
• Probabilistic ML - 15 - Logistic Regression — Bayes' theorem and exponential families provide a probabilistic foundation for logistic regression and softmax functions.
• Hybrid Imbalanced Regression Through Unified Data-Level and Algorithm-Level Balancing — The framework unifies data- and algorithm-level balancing to improve imbalanced regression performance.

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 Computer Vision and Pattern Recognition.