Spiking neural network for time-series prediction


I’ve been trying to use Brian2 to implement a SNN capable of classifying time series. Until now I have used the MNIST dataset (for no special reason). Instead of designing a network with the typical 748 inputs (28x28 pixels per image), I want to design a network that can be fed each column at a time. That is, I define a time interval to encode each column (using time-to-first-spike e.g.). Let’s say this time interval is given the name pattern_dt. In each iteration, I feed the 28 columns (each one of them is encoded independently from the others) to the network resulting in 28*pattern_dt for each image.

My final intention is to classify time series recorded with movement sensors. If I was to feed a network directly from the sensors, I would want to feed it each sample at a time (or a window of samples).

What kind of topology is best to achieve this? From my research, I found that a reservoir would be a good approach but I’m not sure how to implement it.

If anyone can enlighten me in the right way I will much appreciate it.