Interesting. You didn't specify, I'm guessing you did 3x3 convolutions on the spectrographs? Also, how did you choose the convolution size, number of conv/poo

Interesting. You didn't specify, I'm guessing you did 3x3 convolutions on the spectrographs? Also, how did you choose the convolution size, number of conv/pooling layers, etc? Did you consider asymmetric convolution/pooling layers to account for the differences between the frequency and time dimensions? There are a number of interesting directions you could go with that data set. One interesting possibility is to make a convolutional autoencoder, then use that to apply "deep dreaming" filters to music. Another interesting evolution would be to handle the frequency dimension using a 1D convolution, and run a RNN on top of that to deal with time.