alishdipani/nmnist-stdp-spinnaker

A Spiking Neural Network model for Digit Recognition using the N-MNIST dataset.

nmnist-stdp-spinnaker's Issues

Unsupervised learning

Hi, I refer to the training of digit recognition. Is my interpretation correct:

1. We train each output neuron with a specific class. Meaning to say, suppose I have 100 images of class 1 ( 16 x 16 size). I initialise 256 input Poisson neurons with spike rate proportional to the pixel intensity and 1 output neuron, so there are 256 synapses specific to that output neuron. I run the simulation over 300ms x 100 (suppose I expose each image for 300ms) then I extract all the synaptic weights once training is done. Now if I have 100 images of class 2, I initialise a NEW set of 256 input poisson neurons and 1 NEW output neuron and I repeat the workflow, ...

2. In this case, there is no concept of trained model, but rather trained synaptic weights. So each output neuron is characterised by their own 256 synaptic weights. This means that if I have 10 classes, and I use 10 output neurons, I will have 10 sets of 256 synaptic weights and together with the STDP definition, this constitutes a model.

3. To assign a class to that output neuron, all I need is to feed the training images over the 10 sets of 256 synapses and record the spiking activity, albeit at zero learning rate, so weights will always be constant (The ones that are trained), the only change is how the output potential evolves and hits the threshold?

Thank you and have a happy holiday!

The idea behind unsupervised training

Hi, I refer to the training of digit recognition. Is my interpretation correct:

1. We train each output neuron with a specific class. Meaning to say, suppose I have 100 images of class 1 ( 16 x 16 size). I initialise 256 input Poisson neurons with spike rate proportional to the pixel intensity and 1 output neuron, so there are 256 synapses specific to that output neuron. I run the simulation over 300ms x 100 (suppose I expose each image for 300ms) then I extract all the synaptic weights once training is done. Now if I have 100 images of class 2, I initialise a NEW set of 256 input poisson neurons and 1 NEW output neuron and I repeat the workflow, ...

2. In this case, there is no concept of trained model, but rather trained synaptic weights. So each output neuron is characterised by their own 256 synaptic weights. This means that if I have 10 classes, and I use 10 output neurons, I will have 10 sets of 256 synaptic weights and together with the STDP definition, this constitutes a model.

3. To assign a class to that output neuron, all I need is to feed the training images over the 10 sets of 256 synapses and record the spiking activity, albeit at zero learning rate, so weights will always be constant (The ones that are trained), the only change is how the output potential evolves and hits the threshold?

Thank you and have a happy holiday!

Get Accuracy

@alishdipani Hi, I would like to know how to get accuracy from the result. Will the code update in the future? Thanks.