trends.barnas.net

phenomenon · first seen 30 Oct 2026

symmetry breaking in neural network training

Neurons starting from identical or small initializations differentiate and specialize to distinct features during gradient training.

Mentions per week

0119 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Earliest sources

Latest sources

Mentioned together with

gradient flowtwo-layer ReLU network teacher-student model