Work on 2017 W42

US

We're forging ahead primarily in the portion of the project which does not overlap - the multi-spiking neural net and its associated training algorithm. We started to write out the internal state, at first prompting for each weight, to isolate the minimum possible segment. It attempts none of the training, heuristics, or parameter selection which is found in Ghosh-Dastidar and Adeli (2009).

We did develop a small amount of preprocessing, however. We found a reference to a gradient-based preprocessing technique called address-event representation (the actual technique is far more complex, but this was a potentially relevant step). This should allow smaller spikes in the EEG to trigger spikes in the neural net. Since computing a derivative per se is not possible with discrete data, we took the difference between each successive pair of readings and used this as a proxy. These differences seemed (subjectively) to poorly correlate with an instinctive approach at determining where one could say a "spike" occurred. Also, while this may be of no ultimate significance, the differences between successive voltages are not normally distributed.