Lost or Hidden? A Concept-Level Forgetting in Supervised Continual Learning

2026-05-19 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, extended

Summary

A new diagnostic framework, leveraging Sparse Autoencoders (SAEs), has been developed to analyze catastrophic forgetting in continual learning at a fine-grained, concept-level. This framework, proposed by Filus et al., moves beyond traditional task-level performance metrics to investigate how task-specific information evolves within a vision model's representation space. By defining a task-anchored latent feature space where individual SAE latents act as concept proxies, the researchers decompose forgetting into apparent concept deletion, recoverability, and decodability. Their findings, based on experiments with Resnet18 on 2seq-CIFAR10, 2seq-tiny-ImageNet, and 10seq-tiny-ImageNet datasets, indicate that a significant portion of seemingly lost concept-level information can be recovered under a linearity assumption. This suggests that forgetting often stems from changes in representational accessibility rather than complete information erasure, although concept decodability degrades as more tasks are introduced.

Key takeaway

For research scientists developing continual learning models, understanding the nature of forgetting at a concept level is crucial. You should integrate diagnostic frameworks using Sparse Autoencoders to distinguish between true information loss and mere representational drift. Prioritize strategies like Learning without Forgetting (LwF) that maintain linear recoverability of fine-grained concepts, as this preserves model interpretability and stability, even if it means a slight trade-off in overall per-task performance compared to methods like DER++.

Key insights

Catastrophic forgetting in continual learning is often representational misalignment, not complete information erasure.

Principles

• Forgetting can be decomposed into deletion, recoverability, and decodability.
• Linear translation can restore much of the lost concept-level information.
• Continual learning strategies differ in fine-grained information preservation.

Method

Train Sparse Autoencoders (SAEs) on task data to define a fixed, task-anchored latent space. Analyze concept activation dynamics, linear recoverability via translation, and concept-level decodability using linear probes.

In practice

• Use SAEs to diagnose fine-grained forgetting in vision models.
• Employ linear translation to assess concept recoverability.
• Evaluate concept decodability with linear classifiers.

Topics

• Continual Learning
• Catastrophic Forgetting
• Sparse Autoencoders
• Concept-Level Analysis
• Representational Drift

Code references

• aimagelab/mammoth
• KempnerInstitute/overcomplete

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

Related on AIssential

• Forgetting is Not Erasure: Recovering Latent Knowledge via Transport Keys — Catastrophic forgetting often reflects interface drift, not erasure, allowing latent knowledge recovery via alignment.
• 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.
• Catastrophic Forgetting is Low-Rank: A Function-Space Theory for Continual Adaptation — Catastrophic forgetting is a low-rank function-space phenomenon, predictable via a cross-task kernel in the NTK regime.
• Teaching AI to Remember: What I Learned Building a Continual Learning Framework for Drug Discovery — Continual learning models can avoid catastrophic forgetting by integrating new knowledge with awareness of existing representations.
• A Quantitative Characterization of Forgetting in Post-Training — Forgetting in generative models is quantifiable, driven by divergence direction, geometric overlap, and sampling.
• Evaluating the Impact of Task Granularity on Catastrophic Forgetting in Continual Learning — Learning general categories before specific classes may reduce catastrophic forgetting in continual learning systems.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.