Neural Networks Trained on the Same Task Cluster in Matching Weight Structures
After removing hidden-neuron permutations and ReLU rescaling freedoms, independently trained networks fell into reproducible task-dependent weight clusters, with dataset-specific structure ...
