In many papers about artificial spiking neural networks (SNNs), the performance of them is not up to par with traditional ANNs. I have read how some people have converted ANNs to SNNs using various techniques.
There has been work done on using unsupervised learning in SNN to recognise MNIST digits through spike-timing-dependent plasticity (for example, the paper Unsupervised learning of digit recognition using spike-timing-dependent plasticity, by Diehl and Cook, 2015). This form of learning is not possible in traditional ANNs due to their synchronous nature.
I was wondering would it be a good idea to first train an SNN in an unsupervised manner to learn some of the structure of the data. Then convert to a traditional ANN to take advantage of their superior performance with some more training. I can see this being useful for training a network on a sparsely labelled dataset.
I am quite a novice in this area, so I was looking for feedback on any immediate barriers as to why this would not work or if it is even worth doing.
