[Paper Review] Probing Representation Forgetting in Supervised and Unsupervised Continual Learning

MohammadReza Davari, Nader Asadi+3 more|arXiv (Cornell University)|Mar 24, 2022

Domain Adaptation and Few-Shot Learning4 citations

TL;DR

This paper introduces a novel evaluation framework for continual learning (CL) that uses optimal linear probes (LP) to measure representation forgetting, revealing that naive fine-tuning with supervised contrastive learning (SupCon) causes minimal representation forgetting—often outperforming dedicated CL methods. The key contribution is demonstrating that SupCon-based fine-tuning with small memory-based prototype recall enables fast recovery and competitive performance without explicit forgetting control.

ABSTRACT

Continual Learning research typically focuses on tackling the phenomenon of catastrophic forgetting in neural networks. Catastrophic forgetting is associated with an abrupt loss of knowledge previously learned by a model when the task, or more broadly the data distribution, being trained on changes. In supervised learning problems this forgetting, resulting from a change in the model's representation, is typically measured or observed by evaluating the decrease in old task performance. However, a model's representation can change without losing knowledge about prior tasks. In this work we consider the concept of representation forgetting, observed by using the difference in performance of an optimal linear classifier before and after a new task is introduced. Using this tool we revisit a number of standard continual learning benchmarks and observe that, through this lens, model representations trained without any explicit control for forgetting often experience small representation forgetting and can sometimes be comparable to methods which explicitly control for forgetting, especially in longer task sequences. We also show that representation forgetting can lead to new insights on the effect of model capacity and loss function used in continual learning. Based on our results, we show that a simple yet competitive approach is to learn representations continually with standard supervised contrastive learning while constructing prototypes of class samples when queried on old samples.

Motivation & Objective

• To address the limitation of standard accuracy metrics in measuring catastrophic forgetting, which may fail to detect trivial representation changes versus actual knowledge loss.
• To investigate whether representation forgetting—defined as loss of useful features for prior tasks—occurs in standard fine-tuning and self-supervised CL methods.
• To evaluate the impact of model capacity and loss functions on representation forgetting using linear probe-based analysis.
• To propose a simple, memory-efficient method for fast remembering in continual learning that avoids explicit forgetting control mechanisms.
• To demonstrate that representation utility, not just task accuracy, is a critical metric for evaluating continual learning performance.

Proposed method

• Uses optimal linear probes (LP) trained on feature activations from intermediate network layers to evaluate the usefulness of representations for previous tasks.
• Measures representation forgetting as the performance drop of an LP classifier on old task data after training on new tasks.
• Applies this evaluation to standard CL benchmarks (e.g., SplitCIFAR100, ImageNet-Scenes-CUB-Flowers) under both supervised and self-supervised settings.
• Proposes a method combining SupCon fine-tuning with a small memory of class prototypes (NME) to enable fast recovery on old tasks.
• Compares LP-based forgetting to standard accuracy metrics and uses CKA (centered kernel alignment) as a supplementary similarity metric.
• Employs standard ResNet and VGG architectures across multiple task sequences to analyze depth-wise representation changes.

Figure 1 : Performance on ImageNet during the sequence (ImageNet→Scenes→CUB→Flowers) using ResNet18. We observe that although observed accuracy heavily degrades the LP accuracy, in finetuning does not decay as drastically and can rival LP accuracy of methods such as LwF and EWC. Moreover, we observe

Experimental results

Research questions

• RQ1 To what extent does standard accuracy fail to capture true representation forgetting in continual learning?
• RQ2 How does representation forgetting vary across different model architectures and loss functions in continual learning?
• RQ3 Can naive fine-tuning with SupCon or SimCLR achieve better representation retention than dedicated CL methods?
• RQ4 Does increasing model depth or width reduce representation forgetting, and is this effect visible only through LP evaluation?
• RQ5 Can a simple method using SupCon and a small memory of prototypes enable fast remembering without explicit forgetting control?

Key findings

• Representation forgetting under naive fine-tuning is significantly lower than suggested by standard accuracy metrics: only 5.72% LP accuracy drop on ResNet’s final block in a 2-task SplitCIFAR10 setting.
• SupCon fine-tuning with NME memory achieves 53.9% average accuracy on a 10-task SplitCIFAR100 sequence, outperforming LwF (47.7%) and approaching ER-M5 (54.8%) with similar memory size.
• Representation utility remains high in early network layers—LP accuracy increases by 1.08% in ResNet block B-0 and 1.26% in B-1 after task 2, indicating minimal forgetting in low-level features.
• Forgetting is concentrated in deeper layers: LP accuracy drops by 5.72% in ResNet block B-5, but this is still far less catastrophic than the 22% drop in observed task accuracy.
• Wider and deeper models show reduced representation forgetting when evaluated via LP, a trend not clearly visible through standard accuracy metrics.
• The proposed SupCon + NME method enables fast recovery on old tasks without requiring memory access during training, demonstrating strong practical utility.

Figure 3 : 10-Task SplitCIFAR100. We compare observed acc. and linear probe acc. Naively finetuning with CE does poorly if using the observed accuracy. However using the LP based evaluation we observe that performance gap to other methods is lower. Furthermore when finetuning is performed instead wi