I’m studying how Temporal Memory works and I don’t fully understand how exactly basal dendrites learning to predict the next input.
If I understand correctly, in order for the basal dendrites of the winner cell at time step T to learn to predict, they must receive a signal from the winner cells of time step T-1.
This is right?
If this is correct, then how does it happen from a biological stand point? should cells-winners of time step T-1 continue to transmit a signal through their axons at the time step T? After all, it is only at time step T that the winner cell of step T is found, hence their basal dendrites.
Among everyone that gets enought somatic stimulus, only the ones that predicted them fire, that’s why neurons can be connected all they require to be, but only the signals of the ones that learn to predict the input goes to the next “layer”. Hence the next layer can have a very good idea of what is going on because the previous neurons somehow modeled it.
To learn, the biology must have a way for them to know about that signal.
In biology, it’s spike timing dependent plasticity. When a synapse activates (because a presynaptic cell fires), learning doesn’t occur immediately. If the postsynaptic cell fires right after the presynaptic cell fired, then the synapse is reinforced.
The effect of the synapse must linger for that to happen. Not much of the lingering is in the axon itself, I think.
The excitatory signal from a synapse lasts a bit of time, e.g. 50 milliseconds, and other things can linger such as ion concentrations. There are also metabotropic receptors, which can produce effects lasting seconds.
Yeah, so probably not metabotropic receptors. They just make it possible that the brain can do temporal memory with large timesteps. 50 milliseconds is a reasonable duration for a timestep in the brain, because oscillatory cycles have durations in that neighborhood. STDP windows might be more like 10 ms, I forget, but most STDP won’t be for temporal memory since there are a lot of other mechanisms.