Tag
This paper presents a theoretical framework showing that learning in neural networks generates graph symmetries (fibrations and coverings), which enable drastic model compression and improve continual learning performance.