Tag
This paper shows that the geometric symmetry visible from neural network weights depends on the positional encoding and readout observable, and validates this using MLPs trained on 2D signed distance functions with multiple symmetry groups.