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

Researchers at arXiv have identified temporal imbalance as a key factor causing catastrophic forgetting in Class-Incremental Learning (CIL) systems. They propose Temporal-Adjusted Loss (TAL), a new method that uses temporal decay kernels to reweight negative supervision, demonstrating significant improvements in reducing forgetting across multiple CIL benchmarks.

• →Temporal imbalance, where earlier classes receive stronger negative supervision toward the end of training, is identified as a key cause of prediction bias in Class-Incremental Learning.
• →The proposed Temporal-Adjusted Loss (TAL) method uses temporal decay kernels to dynamically reweight negative supervision in cross-entropy loss.
• →TAL degenerates to standard cross-entropy under balanced conditions while effectively mitigating prediction bias under imbalance.
• →Extensive experiments show TAL significantly reduces catastrophic forgetting and improves performance on multiple CIL benchmarks.
• →The research emphasizes the importance of temporal modeling for achieving stable long-term learning in AI systems.