Introduction
In the field of Continual Learning (CL), models are tasked with learning from a stream of data without forgetting previously acquired knowledge—a challenge known as catastrophic forgetting. Traditional approaches often rely on "task boundaries" or supervised labels to manage memory buffers. However, real-world scenarios are frequently Online Task-Free (OTFCL), where data arrives in small batches without explicit indicators of distribution shifts or available labels.
To address this, researchers have introduced Dynamic Cluster Memory (DCM). This non-parametric system organizes memory into evolving clusters based on the data's inherent structure, enabling both generative and discriminative learning from unlabeled, continuous streams.
The Dynamic Cluster Memory (DCM) Framework
The core of the proposed method is a memory system that treats stored samples as a collection of clusters that can grow or shrink dynamically.
1. Knowledge Discrepancy and Expansion
To determine if a new data point represents a novel concept, DCM employs a Knowledge Discrepancy Measure (KDM). Two variants are proposed: DCM-SE (Square Error) and DCM-JS (Jensen-Shannon divergence). The latter uses a forward diffusion process to make raw data distributions mathematically tractable, providing a robust novelty metric. If a sample's discrepancy exceeds a threshold ($\lambda$), a new memory cluster is created.
2. Sample Selection and Pruning
Incoming samples are assigned to the most similar existing cluster prototype. To maintain a fixed memory capacity while preserving diversity, a Memory Pruning Process (MPP) is used. This graph-based evaluation identifies and removes clusters with significant informational overlap, ensuring the memory remains efficient and representative.
3. Integration with Generative Models
DCM is uniquely suited for Denoising Diffusion Probabilistic Models (DDPMs). By providing high-quality "replay" samples from its clusters, it allows generative models to maintain performance across shifting distributions without explicit task labels.
Key Insights and Results
DCM is a "plug-and-play" module that can be integrated into various architectures. Its reliance on cluster prototypes rather than individual samples makes it computationally efficient.
- Generative Excellence: On the Split CIFAR10 dataset, DCM-JS achieved a Frechet Inception Distance (FID) of 76.58, a significant improvement over the previous state-of-the-art (115.38). It also successfully generated high-resolution ($256 \times 256$) images on CelebA-HQ.
- Superior Classification: In supervised settings, DCM-JS outperformed established methods like iCARL and CoPE on Split CIFAR100, reaching 26.27% accuracy.
- Efficiency: The system maintains high performance with fewer stored samples by effectively pruning redundant information.
Conclusion
Dynamic Cluster Memory offers a robust solution for online task-free continual learning. By leveraging non-parametric clustering and diffusion-based distance metrics, it bridges the gap between supervised and unsupervised memory management. Future work may explore more complex graph relations for pruning and extending the framework to other modalities like video and natural language.
Summary updated