On the geometry of feedforward neural network error surfaces
Neural Computation
Artificial Neural Networks: Approximation and Learning Theory
Artificial Neural Networks: Approximation and Learning Theory
Clustering in Weight Space of Feedforward Nets
ICANN 96 Proceedings of the 1996 International Conference on Artificial Neural Networks
Hi-index | 0.00 |
We describe symmetries of feedforward networks in terms of theircorresponding groups, which naturally act on and partition weight space.This leads to an algorithm that generates representative weight vectors ina specific fundamental domain. The closure of this domain turns out to bea manifold with singular points. We derive a canonical metric for themanifold that can be implemented efficiently even for large networks. Oneapplication would be the clustering of resulting weight vectors of anexperiment in order to identify inadequate models or learning methods.