About this document
7768 Heads Collapse Features S by weifeng2829 is a document available to read on EtoBox.
This paper explores the paradox of deep and shallow forgetting in continual learning, revealing that while minimal replay buffers can prevent deep forgetting, larger buffers are necessary to mitigate shallow forgetting. The authors extend the Neural Collapse framework to characterize the geometry of feature representations in continual learning, demonstrating that shallow forgetting arises from classifier misalignment due to under-determined optimization. They propose that understanding the interplay betwee
- Author
- weifeng2829
- Language
- EN