6 replies to this topic

[TheSourceOfX]

Hi everyone. I'm looking for some information. I'm trying to learn more about artificial neural networks which have some form of time dependency. The standard ANN model doesn't include any concept of time, so I'm not sure what to look for or where to look for it when searching for time dependency. Does anyone have any ideas?

[WingedPanther]

In what sense do you want a time dependency?

[TheSourceOfX]

Really I want any sort of time dependency I can find. Ideally I would like to see something where the connections between neurons have lengths and the signals have velocities. However, anything type of neural network involving time in some way would be lovely.

[zoranh]

So you want to have a neural network which has signal latency inside it?

Such network would converge to its ideal (zero latency) counterpart's state anyway. So what is the purpose of having latency?

Do you expect it to show some specific curve on the output or what? If so, then you just need to add another input accepting discrete time, and then to teach the network how to produce output based on time as well as other inputs.

[Sinipull]

You could try making a step based neural network. Every signal takes 1 unit of time to reach next neuron, so output doesn't react instantly to the input, but takes certain amount of time, or "turns" in your program. There's no need to put the thread to sleep, this will not make any difference, as said before. Step based algorithm will also give your network limits, so it won't grow too big if you use evolution and natural selection to grow it. I also believe that's how actual brains work - larger ones have much bigger reaction times, small ones can react almost instantly.

[TheSourceOfX]

My actual thought was that each neuron would employ a capacitor like device, pretty much like normal, but that each capacitor drained slowly over time. That way the order in which inputs are received matters. Also, a networks such as this probably wouldn't do much if it was strictly feedforward. But do you guys know of any articles or papers discussing any form of neural network implementation that includes time?

[Momerath]

4.1 Integrate-and-fire model
http://citeseerx.ist...p=rep1&type=pdf
Computing with the Leaky Integrate-and-Fire Neuron: Logarithmic Comput... - Microsoft Academic Search

This the type of thing you are looking for?