Reproduce “Unsupervised Continual Learning with Growing Self-Organizing Maps and Synthetic Replay” on OpenResearch

About this paper

Pujan Thapa, Alexander Ororbia, Travis Desell

The authors introduce a growing self-organizing map that learns from streams without storing raw examples. It expands when patterns are poorly represented, stores per-unit distribution statistics, and samples synthetic replay from them. On Split CIFAR-10, combined error-and-surprisal growth reports accuracy of 50.11 ± 0.89%, versus 44.87 ± 1.03% with error-only growth; these are protocol-specific results, not evidence of statistical significance.

Key insights

• Combining adaptive topological learning via GSOMs with statistical, exemplar-free generative replay effectively mitigates catastrophic forgetting in unsupervised and task-free continual learning.
• A surprisal-driven growth criterion, based on Mahalanobis distance, allows the model to autonomously detect and adapt to novel data patterns, leading to more efficient resource allocation.
• Operating GSOMs on high-quality latent representations from powerful pretrained encoders (e.g., CLIP) significantly enhances performance on complex, high-dimensional data, often outperforming supervised baselines.

Summary by alphaXiv.