Temporal Imbalance of Positive and Negative Supervision in Class-Incremental Learning

2026-03-01

AI Summary

This work addresses the issue of prediction bias and catastrophic forgetting of early classes in class-incremental learning, caused by excessively strong negative supervision during later training phases. The study is the first to formally characterize this temporal supervision imbalance and introduces Temporal Adjustment Loss (TAL), which dynamically reweights the negative supervision terms in the cross-entropy loss via a time-decay kernel to modulate supervision strength over time. Theoretical analysis elucidates the degeneration mechanism inherent in conventional approaches, while extensive experiments demonstrate that TAL substantially mitigates forgetting and enhances performance across multiple class-incremental learning benchmarks, thereby underscoring the critical role of temporal modeling in enabling stable long-term learning.

Technology Category

• Machine Learning: Life-Long and Continual Learning
• Natural Language Processing: Learning & Optimization for NLP
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

With the widespread adoption of deep learning in visual tasks, Class-Incremental Learning (CIL) has become an important paradigm for handling dynamically evolving data distributions. However, CIL faces the core challenge of catastrophic forgetting, often manifested as a prediction bias toward new classes. Existing methods mainly attribute this bias to intra-task class imbalance and focus on corrections at the classifier head. In this paper, we highlight an overlooked factor -- temporal imbalance -- as a key cause of this bias. Earlier classes receive stronger negative supervision toward the end of training, leading to asymmetric precision and recall. We establish a temporal supervision model, formally define temporal imbalance, and propose Temporal-Adjusted Loss (TAL), which uses a temporal decay kernel to construct a supervision strength vector and dynamically reweight the negative supervision in cross-entropy loss. Theoretical analysis shows that TAL degenerates to standard cross-entropy under balanced conditions and effectively mitigates prediction bias under imbalance. Extensive experiments demonstrate that TAL significantly reduces forgetting and improves performance on multiple CIL benchmarks, underscoring the importance of temporal modeling for stable long-term learning.

Problem

• Class-Incremental Learning
• Temporal Imbalance
• Catastrophic Forgetting
• Prediction Bias
• Negative Supervision

Innovation

• Temporal Imbalance
• Class-Incremental Learning
• Negative Supervision
• Temporal-Adjusted Loss
• Catastrophic Forgetting