Shapley Neuron Valuation Improves Continual Learning Without Memory Buffers
A recent paper in machine learning presents Shapley Neuron Valuation (SNV), a novel framework that assesses the significance of neurons in continual learning through the lens of cooperative game theory. By selectively freezing crucial neurons and allowing others to remain adaptable, SNV facilitates continual learning without the need for additional architecture. Tests conducted on ImageNet-1k indicate that SNV surpasses current buffer-free approaches, achieving a +2.88% increase in accuracy for class incremental learning and a +6.46% boost for task incremental learning. This research is available on arXiv within the computer science and machine learning sections.
Key facts
- SNV uses cooperative game theory to quantify neuron importance
- Selectively freezes important neurons to prevent catastrophic forgetting
- Buffer-free method, no memory replay or architecture expansion needed
- Tested on ImageNet-1k dataset
- Improves class incremental learning accuracy by +2.88%
- Improves task incremental learning accuracy by +6.46%
- Published on arXiv
- Addresses catastrophic forgetting in neural networks
Entities
Institutions
- arXiv