What are there the best ways to leverage the unique "geometrical" constraints of spectro-temporal signal representations (architecture, filter shapes, data augmentation, etc.)? For example, a learned feature in a spectrogram should intuitively be horizontally (temporally) invariant since the window of time when it was captured is arbitrary. In contrast, the specific vertical position of a feature is extremely important since the precise frequency at which it occurred is significant. My research group is trying do design a convolutional neural network (using keras/tensorflow) that takes spectrograms representing neural data and predicts a particular disease biomarker. All of our trained models so far severely overfit and we have somewhat limited data.
Just for further background:
We decided to use CNNs because the visual representations of spectral data are known to be useful for our task. Even though our data set is not huge, we were hoping that CNNs could be more interpretable in terms of exactly what spectral features are being learned than other ML methods like SVMs or LSTMs. The models have been regressors so far, but we are open to switching to classifiers if they perform better in the end.