Introduction
Class-Incremental Learning (CIL) remains a significant challenge in computer vision. When a model trained on a set of initial classes is subsequently tasked with learning new ones, it typically suffers from "catastrophic forgetting"—the abrupt loss of performance on the original data. While "replay" strategies involving a small memory buffer of old samples help mitigate this, they often introduce a severe data imbalance that standard training methods struggle to resolve.
The Stability-Plasticity Conflict
To combat forgetting, current state-of-the-art methods rely on distillation losses that force the model's internal feature representations (embeddings) for old classes to remain as close as possible to their original values. However, this "point-wise" distillation is often too rigid. As a model incorporates new knowledge, its feature space naturally needs to reorganize. Forcing exact coordinate matches creates a stability-plasticity conflict: the model becomes either too rigid to learn new classes effectively or too flexible, leading to the erasure of old knowledge.
Triplet Contrastive-relationship Preserving (TCP)
The authors propose a shift in perspective: instead of preserving exact feature values, we should preserve the relative relationships between features. This is implemented via the Triplet Contrastive-relationship Preserving (TCP) loss.
TCP ensures that if Image A was more similar to Image B than to Image C in a previous phase, that ranking is maintained in the current phase. This allows the entire feature space to "drift" and evolve globally to accommodate new classes while keeping the discriminative structure of the old classes intact. An oracle experiment conducted by the researchers revealed that while feature values change significantly, over 77% of these triplet relationships remain stable in an ideal model, justifying this relational approach.
Asymmetric Augmented Contrastive Learning (A2CL)
To complement TCP, the framework employs Asymmetric Augmented Contrastive Learning (A2CL). Standard Supervised Contrastive Learning (SCL) often fails under the extreme data imbalance found in CIL. A2CL addresses this by applying different augmentation strategies:
- Old Data: Uses "Strong Augmentation" for all views to prevent overfitting on the limited samples in the memory buffer.
- New Data: Uses a mix of "Weak" and "Strong" augmentations to ensure the model learns robust boundaries that can withstand future drift.
Key Insights and Results
The core insight of this work is that the structure of the feature space is more critical for recognition than specific coordinate values. By relaxing the distillation constraint, the model gains the plasticity required to integrate new classes without sacrificing the stability of old ones.
The method was evaluated on CIFAR-100, ImageNet-100, and ImageNet-1000. Key results include:
- Performance Gains: TCP achieved a 1.26% improvement on CIFAR-100 and a 0.95% improvement on ImageNet over previous state-of-the-art methods.
- Memory Efficiency: In highly constrained scenarios (e.g., only 10 images per class), the performance gap widened to 3.2%, demonstrating superior robustness.
- Versatility: TCP can be integrated into other CIL frameworks (like RMM) as a "plug-and-play" component to enhance their performance.
Conclusion
By moving from point-wise distillation to relationship-preserving distillation, the TCP framework offers a more flexible and effective solution for Class-Incremental Learning. While the current implementation relies on a small memory buffer, future research could explore applying these relational principles to data-free incremental learning or unsupervised settings.
Summary updated