I have been reading latley on SNN and decided to try and implement myself some simple simulation. So I wrote a simple MATLAB simulation using simple LIF current based model and I try using STDP to classify unsupervised MNIST data such as here: https://www.frontiersin.org/articles/10.3389/fncom.2015.00099/full#
(So, images converts to poisson spike train by intensity)
I'm using simple window based STDP updates and most of the times I'm getting that a single same neuron in the hidden layer spikes for any input sample that I feed in (i.e. only neuron #50 spikes for any given input).
I would really appreciate if anyone can clarify the following for me:
Is there a way to normalize/regularize the weights updates such that this neuron dominance won't happen?
Any general tips on how to monitor/verify the STDP convergence? max/min of weights? positive STDP update larger than negative STDP update?
Again, would really appreciate any help. Thanks