Fig. 4: The distribution of weights. | Nature Communications

Fig. 4: The distribution of weights.

From: Machine learning in spectral domain

Fig. 4

Distribution of the weights of a perceptron. The red line follows the spectral training limited the N1 + N2 eigenvalues. The black line follows the training in direct space. Here, N1 × N2 parameters are adjusted in the space of the nodes. The distribution is very similar, but the spectral learning employs about 10% of the parameters used in direct space. The distributions obtained when forcing the training in direct space to operate on a subset of N1 + N2 weights are very different from the one displayed (for every choice of the randomly selected family of weights to be trained).

Back to article page