Monday, October 2, 2017

Neurons, Synapses, Plasticity, Etc.

A Funny Thing Happened

Has anyone noticed that I wrote 2 articles (see links below) on unsupervised learning in spiking neural networks and not once did I mention the properties of neurons and their synapses? I never said a word about spike timing dependent plasticity, synaptic strength or signal integration or anything of the sort. Guess what? It does not matter how neurons are implemented biologically or artificially. What matters is what they do, i.e., what function they perform as a whole and how they acquire their connections. Pattern neurons, for example, do one thing while they are learning and a different thing when they are no longer learning.

Biologically speaking, pattern neurons receive precisely timed synchronization signals from the hippocampus. There is also complex neural circuitry that comes into play during pattern pruning. It gets much more complicated in sequence memory. But this is all irrelevant to the principles we need to understand.

I always smile when I see people agonizing over what properties neurons should have. This is not what is important. The bigger principles are what we should be focusing on. Implementation is just engineering. It can be done any number of ways, in hardware or software. Some of you may be surprised to learn that I don't use synaptic strength properties in my own software experiments.

Something to think about.

See Also:

Fast Unsupervised Pattern Learning Using Spike Timing
Fast Unsupervised Sequence Learning Using Spike Timing

No comments: