Hopfield/Brody mouse mus silicium

Contents | Results | Description

Results

The error scores (as defined on this page) for my winning entry to competition B were 0.253 (without random disruption to the weights) and 0.259 (with random 50%/150% "brain damage" applied to weights after training).

My versions of figures 1 and 2 from Hopfield and Brody's paper are below:


Figure 1: Output of the network in response to some input waveforms after training on a single instance of the word "one" and 9 other random patterns. Each panel shows the waveform in the top left, the output of the "one"-recognising gamma neuron for 10 trials, and the corresponding peristimulus time histogram. The waveforms used were a) "one" (speaker a); b) "one" (speaker b); (c) "one"+tone (speaker a); (d) reversed "one2 (speaker a); (e) "three" (speaker b).


Figure 2: Performance of the network on the test set. The trained network was tested on 500 sampled words; 50 of each digit "zero"-"nine". The production of two or more spikes from the "one"-trained gamma neuron was considered to denote recognition. Each bar shows the percentage of samples which were recognised as "one", for each digit. The triangles show the average recognition rate for the five different speakers used in the test set.

[Contents]