Shapley Neuron Values for Continual Learning: Which Neurons Matter Most?

2026-05-15

AI Summary

This work addresses catastrophic forgetting in continual learning by proposing a Shapley value–based neuron valuation framework grounded in cooperative game theory. It introduces, for the first time, the use of Shapley values to quantify the importance of individual neurons to previously learned tasks and selectively freezes the most critical ones to preserve knowledge. Notably, the method achieves this without relying on replay buffers or expanding the model architecture. Experimental results on ImageNet-1k demonstrate that the approach significantly outperforms existing buffer-free strategies, yielding absolute accuracy gains of 2.88% and 6.46% under class-incremental and task-incremental settings, respectively, thereby confirming its effectiveness and novelty.

Abstract

Continual learning enables neural networks to learn tasks sequentially without forgetting previously acquired knowledge. However, neural networks suffer from catastrophic forgetting, where learning new tasks degrades performance on earlier ones. We address this problem with Shapley Neuron Valuation (SNV), a principled framework that quantifies Neuron importance in continual learning, grounded in cooperative game theory. SNV selectively freezes important Neurons while keeping others plastic, enabling buffer-free continual learning without expanding architecture. Experiments on ImageNet-1k show that SNV consistently outperforms existing buffer-free methods. In particular, SNV improves accuracy by +2.88% in the class incremental learning and +6.46% in the task incremental learning scenarios compared to the second baseline.

Problem

• continual learning
• catastrophic forgetting
• neuron importance
• class incremental learning
• task incremental learning

Innovation

• Shapley Neuron Valuation
• Continual Learning
• Catastrophic Forgetting
• Neuron Importance
• Buffer-free