Kernel Variation

Much of the information about spectra is contained in the relative distance between peaks as well as the peak heights. Our current kernel implementation is constructed to scan across the channels (essentially left to right) and does not allow building of relationships across different regions of the spectra/features.

There are several kernel parameters which can be varied across channels in a layer to introduce more complexity to the feature definitions within a layer and the forward feature relationships across layers. We will experiment variation on kernel parameters and model structure to test for improvement on model prediction accuracy. Of most interest here is perhaps variation in the dilation parameter which introduces spacing between kernel components.

Motivating example: Pairs of peaks at a certain spacing and relative intensity are a pretty powerful feature complement to individual peaks. (take the case of overlapping peaks for different elements) Consider this further in a layered context where peak pairs, at a certain spacing, further refine the combination uniqueness.

Implementation: It seems pytorch introduces spacing between each kernel element equal to dilation so a first attempt would just use a kernel of size=2 and a tuple of various dilation/spacing values. The would be added to the existing "single peak" filter. Stride can still be 1 and no padding needed since edges form the boundary of where "pairs" would exist.