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

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.

Today's AI systems typically learn everything at once from a fixed dataset. But in the real world, new information arrives over time; a medical imaging system, for example, must recognize new diseases without forgetting how to detect earlier ones. Standard neural networks fail at this: learning something new erases what they already knew, a problem called catastrophic forgetting. Our method, Shapley Neuron Valuation, fairly divides credit among members of a team. We use it to measure how much each neuron in a network contributes to what the system has already learned. Neurons that matter most are protected from change; the rest remain free to learn new things. Unlike many competing approaches, ours requires no stored examples from previous tasks and no additional network capacity, making it more practical and more privacy-friendly. On ImageNet, a benchmark containing one thousand visual categories learned in sequence, our method outperforms existing approaches by a meaningful margin. This brings us closer to AI systems that can continuously absorb new knowledge the way people do, an important capability for applications like robotics, personalized medicine, and any setting where the world doesn't stop changing after training ends.