Universal Incremental Learning: Mitigating Confusion from Inter- and Intra-task Distribution Randomness

Sheng Luo, Yi Zhou, Tao ZhouarXiv 2025 · DOI · arXiv

Introduction

Incremental Learning (IL) aims to enable models to learn from a continuous stream of data without forgetting previous knowledge. However, traditional IL paradigms often rely on rigid assumptions. Class-IL assumes only new classes arrive, while Domain-IL assumes only new environments appear. Even Versatile IL (VIL), which combines both, typically assumes a fixed and known scale for these increments.

In real-world scenarios, such as autonomous driving or robotics, these assumptions rarely hold. A model might encounter a varying number of new classes in one task and a completely new domain with familiar classes in the next. This paper introduces Universal Incremental Learning (UIL), a framework that removes these constraints, requiring models to handle unpredictable sequences of tasks where both the type and scale of increments are unknown.

Comparison of different Incremental Learning paradigms: CIL, DIL, VIL, and the proposed UIL.

Challenges in Universal Incremental Learning

The transition to UIL introduces two primary sources of "confusion":

  1. Inter-task Distribution Randomness: The unpredictable nature of task increments (e.g., switching between class-heavy and domain-heavy tasks) leads to non-deterministic predictions and low model confidence.
  2. Intra-task Distribution Randomness: Real-world data often exhibits a long-tail distribution. Varying sample counts per class lead to asymmetrical optimization, where frequent "head" classes dominate the learning process at the expense of rare "tail" classes.

The MiCo Framework

To address these challenges, the authors propose MiCo (Mitigating Confusion), a framework designed for stable and balanced learning in UIL environments. MiCo consists of three core components:

Multi-Objective Learning

MiCo integrates standard Cross-Entropy (CE) loss with Entropy Minimization (EM). While CE focuses on accuracy, EM encourages the model to produce "sharp," high-confidence predictions. The authors found that reducing prediction distribution entropy helps the model better distinguish between conflicting knowledge across diverse task types.

Direction Recalibration Module

Optimizing for both accuracy and confidence can lead to gradient conflicts. This module monitors the Gradient Cosine Similarity between the CE and EM objectives. If gradients are antagonistic, it adjusts the update direction to ensure harmonious learning across both goals.

Magnitude Recalibration Module

To combat intra-task imbalance, MiCo rescales gradient magnitudes based on class frequency. This prevents "head" classes from overwhelming the optimization, ensuring that "tail" classes receive sufficient gradient signal for fair representation.

The relationship between gradient magnitude, class, and accuracy, highlighting the need for magnitude recalibration.

Experimental Evaluation

The MiCo framework was evaluated on three adapted benchmarks: iDigits, CORe50, and DomainNet, modified to simulate UIL by introducing randomness in class and domain increments.

Key Results

MiCo consistently outperformed state-of-the-art (SOTA) methods, including ICON, DualPrompt, and CODA-Prompt:

  • CORe50: Achieved a significant +6.4% improvement in mean accuracy over previous best methods.
  • DomainNet: Outperformed SOTA by +2.7%, demonstrating robustness in large-scale, multi-domain scenarios.
  • iDigits: Reached 77.1% mean accuracy, a +2.2% gain.

Importantly, MiCo maintained competitive forgetting rates, suggesting that its focus on prediction confidence does not lead to catastrophic interference with prior knowledge.

Conclusion

Universal Incremental Learning represents a significant step toward deploying AI in dynamic, real-world environments. By identifying and mitigating the confusion caused by inter- and intra-task randomness through gradient and entropy recalibration, the MiCo framework provides a robust foundation for future research in unconstrained incremental learning.

Summary updated

BibTeX
@misc{luo2025universal,
  author = {Sheng Luo and Yi Zhou and Tao Zhou},
  title = {{Universal Incremental Learning: Mitigating Confusion from Inter- and Intra-task Distribution Randomness}},
  year = {2025},
  eprint = {2503.07035},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.07035},
}
Figure