Continual Learning with Weight Interpolation

TL;DR

The paper addresses catastrophic forgetting in continual learning by introducing CLeWI, a weight-interpolation technique that merges previously learned weights with those trained on a new task through permutation-aligned interpolation after each task. By leveraging a memory buffer and aligning activations, CLeWI helps consolidate knowledge across tasks and can be combined with existing rehearsal-based methods to improve accuracy and reduce forgetting, while offering a tunable stability-plasticity trade-off via the interpolation parameter $\alpha$. The authors provide theoretical motivation based on joint loss and forgetting, outline a practical algorithm with permutation-based alignment and BN normalization, and validate the approach on standard benchmarks, showing consistent gains over strong baselines and demonstrating how architecture width and memory constraints influence effectiveness. The method is proposal as a low-cost, plug-in enhancement for continual learning systems, with potential applicability to broader settings beyond image classification and clear paths for future improvements such as selective interpolation and advanced task settings.

Abstract

Continual learning poses a fundamental challenge for modern machine learning systems, requiring models to adapt to new tasks while retaining knowledge from previous ones. Addressing this challenge necessitates the development of efficient algorithms capable of learning from data streams and accumulating knowledge over time. This paper proposes a novel approach to continual learning utilizing the weight consolidation method. Our method, a simple yet powerful technique, enhances robustness against catastrophic forgetting by interpolating between old and new model weights after each novel task, effectively merging two models to facilitate exploration of local minima emerging after arrival of new concepts. Moreover, we demonstrate that our approach can complement existing rehearsal-based replay approaches, improving their accuracy and further mitigating the forgetting phenomenon. Additionally, our method provides an intuitive mechanism for controlling the stability-plasticity trade-off. Experimental results showcase the significant performance enhancement to state-of-the-art experience replay algorithms the proposed weight consolidation approach offers. Our algorithm can be downloaded from https://github.com/jedrzejkozal/weight-interpolation-cl.

Figures (6)

• Figure 1: Continual learning with weight interpolation.
• Figure 2: The effect of the $\alpha$ parameter (\ref{['eq:interpolation']}) on the test set accuracy for all tasks. Interpolation with smaller values of $\alpha$ allows for obtaining weights that are closer in loss landscape to the current task, while increasing $\alpha$ means more weights are carried over from previous tasks.
• Figure 3: Impact of increasing the network width on the accuracy barrier and continual learning performance. (Left) the interpolation plot for the WideResNet with width multiplier $=4$. (Right) test accuracy for split-Cifar100 benchmark as a function of ResNet width.
• Figure 4: Test accuracy in the function of interpolation alpha for CLeWI with no replay
• Figure 5: Interpolation plots for other forms of rehearsal