|
Post by Alireza on May 9, 2016 7:23:48 GMT
This has been of my interest in recent two years, but haven't got an idea so far. Here is a nice paper www.sciencedirect.com/science/article/pii/S0896627314002566Sompolinsky's works are not easy to understand but they deserve to be read carefully. The idea is how to change afferent synaptic weights to a neuron to make a desired map of input-output pattern of spikes. It is a step forward to the "Tempotron" which has been formerly put forward by Gutig and Sompolinsky, and is related to the stream of "learning spatiotemporal patterns" which is being actively followed by Dean Buonomano.
|
|
|
Post by Alireza on May 9, 2016 7:40:32 GMT
I have got a few naive questions when reading this:
- Is there any degeneracy in the form of multiple pattern of weights which lead a single input-output pattern? - Isn't it more realistic to consider a decisive window around each output spike to teach the system and ignore all the rest input spikes? - What is the advantage of inserting a "reservoir" between input and output layers?
|
|