[Paper Review] Continual Feature Selection: Spurious Features in Continual Learning

Timothée Lesort|arXiv (Cornell University)|Mar 2, 2022

Domain Adaptation and Few-Shot Learning4 citations

TL;DR

This paper investigates spurious and local spurious features in continual learning, demonstrating that models often rely on non-generalizable features due to data distribution shifts and limited data access. It shows that performance drops are not solely due to catastrophic forgetting but also stem from overfitting to local spurious features, and proposes replay-based methods to mitigate this issue.

ABSTRACT

Continual Learning (CL) is the research field addressing learning without forgetting when the data distribution is not static. This paper studies spurious features' influence on continual learning algorithms. We show that continual learning algorithms solve tasks by selecting features that are not generalizable. Our experiments highlight that continual learning algorithms face two related problems: (1) spurious features and (2) local spurious features. The first one is due to a covariate shift between training and testing data, while the second is due to the limited access to data at each training step. We study (1) through a consistent set of continual learning experiments varying spurious correlation amount and data distribution support. We show that (2) is a major cause of performance decrease in continual learning along with catastrophic forgetting. This paper presents a different way of understanding performance decrease in continual learning by highlighting the influence of (local) spurious features in algorithms capabilities.

Motivation & Objective

• To investigate the impact of spurious features—correlations between features and labels that do not generalize—on continual learning algorithms.
• To identify and analyze a new problem in continual learning: local spurious features, which arise due to incomplete data availability at each training step.
• To evaluate whether existing continual learning methods, particularly rehearsal, can effectively mitigate the negative effects of spurious and local spurious features.
• To propose a modified Out-of-Distribution (OOD) generalization method tailored for continual learning and assess its performance on a new benchmark.
• To reframe performance degradation in continual learning as stemming from feature selection errors, not only from catastrophic forgetting.

Proposed method

• Proposes a new continual learning benchmark, SpuriousCIFAR2, inspired by colored MNIST, to systematically study spurious correlations with controlled data distribution shifts.
• Employs a binary classification setup with color as a spurious feature to simulate covariate shift between training and test data distributions.
• Introduces a modified OOD generalization method adapted for continual learning, aiming to learn invariant, non-spurious features.
• Uses both single-head and multi-head classifier setups to compare performance and isolate the impact of feature selection on generalization.
• Employs linear and mean-layer classifiers with and without weight normalization to assess whether feature selection bias or norm imbalance explains performance gaps.
• Applies rehearsal-based continual learning methods to simulate an i.i.d. data distribution across tasks, thereby reducing the risk of learning local spurious features.

Experimental results

Research questions

• RQ1 To what extent do continual learning algorithms overfit to spurious features due to covariate shift between training and test data?
• RQ2 How does the presence of local spurious features—features that correlate with labels within a task but not across the full scenario—affect continual learning performance?
• RQ3 Can standard continual learning methods like rehearsal effectively mitigate the negative impact of spurious and local spurious features?
• RQ4 Is performance degradation in continual learning primarily due to catastrophic forgetting, or do poor feature selection decisions (e.g., relying on local spurious features) play a significant role?
• RQ5 Can OOD generalization techniques be adapted and effectively applied in continual learning to improve robustness to spurious features?

Key findings

• Continual learning models frequently rely on spurious features that correlate with labels in training data but not in test data, leading to poor generalization.
• The performance gap between single-head and multi-head classifiers is significantly larger on linear layers than on mean layers, indicating that feature selection bias—not model architecture—drives performance drops.
• Pre-trained models do not mitigate the problem of local spurious features, showing that strong feature extractors alone are insufficient for robust continual learning.
• Replay methods effectively reduce the risk of learning local spurious features by simulating an i.i.d. data distribution across tasks.
• Performance degradation in continual learning is not solely due to catastrophic forgetting but is substantially exacerbated by overfitting to local spurious features.
• The study identifies local spurious features as a core challenge in continual learning, on par with catastrophic forgetting, and calls for new methods that prioritize feature selection over mere memory preservation.