Weighting of synaptic responses

Josh Reich rogan at auswired.net
Thu Aug 29 00:43:36 EST 1996

Matt Jones <jonesmat at ohsu.edu> wrote:
>This brings me to another question about "standard neural net modelling".
>How is timing incorporated into most models? From the basic course I took
>many years ago, I remember that the entire network would be stepped
>simultaneously, all the inputs and their weighting would be calculated to
>produce the output of each neuron, and then the whole network would be
>stepped again. If this is still the way most of the modelling is being
>done, its pretty different from the situation in the brain (not wrong,
>just different). Actually, its kind of like what would happen if you
>considered the biological activity in time units that were really long
>compared to the durations of most synaptic events (time windows of 100s
>of milliseconds). As if you passed the biological system through a 5 Hz
>low pass filter.  You filter out the spikes, and instead of rate coding
>(or whatever the latest theory is) you get outputs that are graded
>according to the input strength. Makes you wonder why real neurons go
>through all the trouble of firing those little spikes in the first place
Well, as you say many models are completely atemporal, but when
i write a neural network that requires some concept of "time"
i use one of a few methods that i have come up with.
One involves neurons which have the ability to keep track of
their last output, and thus they can find dx/dt 's and respond
accordingly. Another method i use is to set up what in 
computing terms is known as a stream, whereby an input must 
propogate through a series of nodes, with each node performing
as a "standard" neuron, with the added property of adding an
artificial latency to signal propogation. In that way a simple
3 layer neural net will take 3 time cycles to calculate, and so 
Please note most of my algorithms are things i have just come
up with to solve problems i come across. To get a better answer
there are books such as "Computation and The Brain" by Patrica
Churchland, which goes into some detail of these sorts of 
things. I am yet to finish reading the book, but when i am, i 
hope i will be enligtened a bit more to the way that real 
computational neuroscientists make neural networks.



More information about the Neur-sci mailing list