First Layer Map

Two Kohonen maps are used to classify units. The first has 49 nodes laid out in a 7 by 7 grid. The nodes are intialised by perturbing a mean of all the inputs by a random proportion of the standard deviation.

Dot-product similarity function is used with some time and frequency shifting. The input to the net consists of the representations of the extracted units. The output is a series of activation maps - 49 values for each unit representing the similarity between that unit and each of the nodes in the map. This can be represented pictorially as above. The yellower areas show which part of the map is most highly activated by (most similar to) the unit.