Leveraging Complementary Embeddings for Replay Selection in Continual Learning with Small Buffers

2026-04-09 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A new method called Multiple Embedding Replay Selection (MERS) addresses catastrophic forgetting in Continual Learning (CL) by improving sample selection for replay buffers, especially under severe memory constraints. MERS integrates both supervised and self-supervised embeddings using a graph-based approach, departing from most existing methods that rely solely on supervised objectives. This strategy leverages class-agnostic, self-supervised representations that often contain rich, class-relevant semantics. Empirical results demonstrate that MERS consistently outperforms state-of-the-art selection strategies across various continual learning algorithms, showing significant gains in low-memory regimes. On CIFAR-100 and TinyImageNet, MERS enhances performance over single-embedding baselines without increasing model parameters or replay volume, positioning it as a practical, drop-in improvement for replay-based CL.

Key takeaway

For research scientists developing continual learning systems with limited memory, MERS offers a significant performance boost by intelligently selecting replay samples. You should consider integrating MERS into your existing replay-based CL algorithms, particularly when working with small buffer sizes, as it provides consistent improvements without increasing model complexity or memory footprint. This approach can help mitigate catastrophic forgetting more effectively.

Key insights

Integrating supervised and self-supervised embeddings via a graph-based approach improves continual learning replay selection.

Principles

• Self-supervised embeddings encode rich, class-relevant semantics.
• Replay selection critically impacts CL performance under memory constraints.

Method

MERS replaces the buffer selection module with a graph-based approach that integrates both supervised and self-supervised embeddings to select samples for replay.

In practice

• Improve CL performance in low-memory regimes.
• Enhance replay-based CL without adding model parameters.

Topics

• Continual Learning
• Catastrophic Forgetting
• Replay Selection
• Self-Supervised Embeddings
• Supervised Embeddings

Best for: Computer Vision Engineer, Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Not Just After One: Sleep-Inspired Replay Prevents Catastrophic Forgetting After Sequential Tasks — Sleep-inspired replay after sequential training prevents catastrophic forgetting in neural networks, enabling continuous learning.
• FlashbackCL: Mitigating Temporal Forgetting in Federated Learning — FlashbackCL mitigates temporal forgetting in Federated Learning by combining decayed label counts, class-balanced replay, and server-side coreset curation.
• Sparsity, Superposition, and Forgetting: A Mechanistic Study of Representation Retention in Continual Learning — The study reveals that superposition's impact on forgetting in CL is nuanced, interacting with representation strength and capacity.
• Continual Fine-Tuning with Provably Accurate and Parameter-Free Task Retrieval — Proteus offers provably accurate, parameter-free task retrieval for continual fine-tuning by leveraging distinct representation signatures.
• Continual Learning in Modern Hopfield Networks with an Application to Diffusion Models — High-energy, outlier-like samples are more prone to forgetting in continual learning for generative models, and replay is most effective for them.
• Re-Evaluating Continual Learning with Few-Shot Adaptation — Few-shot evaluation offers a more comprehensive assessment of continual learning model stability and plasticity than traditional 0-shot methods.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.