Temporal Imbalance of Positive and Negative Supervision in Class-Incremental Learning
arXiv:2603.02280v1 Announce Type: new 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
arXiv:2603.02280v1 Announce Type: new 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.
Executive Summary
This article presents an innovative approach to addressing the challenge of catastrophic forgetting in Class-Incremental Learning (CIL) by introducing the concept of temporal imbalance as a contributing factor to prediction bias. The authors propose Temporal-Adjusted Loss (TAL), a novel method that dynamically reweights negative supervision in cross-entropy loss using a temporal decay kernel. Experiments show that TAL significantly reduces forgetting and improves performance on multiple CIL benchmarks. The authors' focus on temporal imbalance highlights the importance of considering the temporal aspect of learning in CIL, and the proposed method offers a promising solution to the problem of catastrophic forgetting. The findings have significant implications for the development of stable and long-term learning models in visual tasks.
Key Points
- ▸ Temporal imbalance is identified as a key factor contributing to prediction bias in CIL.
- ▸ Temporal-Adjusted Loss (TAL) is proposed as a novel method to address the issue of catastrophic forgetting.
- ▸ TAL dynamically reweights negative supervision in cross-entropy loss using a temporal decay kernel.
Merits
Strength
The article presents a novel and theoretically sound approach to addressing the challenge of catastrophic forgetting in CIL.
Strength
The proposed Temporal-Adjusted Loss (TAL) method offers a promising solution to the problem of catastrophic forgetting.
Strength
The authors' focus on temporal imbalance highlights the importance of considering the temporal aspect of learning in CIL.
Demerits
Limitation
The article assumes a fixed temporal decay kernel, which may not generalize to all scenarios.
Limitation
The experimental evaluation is limited to a small set of CIL benchmarks.
Expert Commentary
This article makes a significant contribution to the field of CIL by introducing the concept of temporal imbalance and proposing a novel method to address the problem of catastrophic forgetting. The proposed Temporal-Adjusted Loss (TAL) method offers a promising solution to this challenge and has significant implications for the development of stable and long-term learning models in visual tasks. However, the article assumes a fixed temporal decay kernel, which may not generalize to all scenarios, and the experimental evaluation is limited to a small set of CIL benchmarks.
Recommendations
- ✓ To further validate the proposed method, the authors should conduct experiments on a larger set of CIL benchmarks and explore the use of adaptive temporal decay kernels.
- ✓ The proposed Temporal-Adjusted Loss (TAL) method should be integrated into existing CIL frameworks to facilitate its adoption in real-world applications.