I am trying to set up an experiment where a HTM model learns a complex mapping between an input sequence and an output sequence, for example, learning to predict the spike train of an Izhikevich neuron based on its input current.
Is this currently achievable? If so, tips towards the realisation of this would be greatly appreciated. TIA!
Hi.
AFAIK HTM makes next time step prediction on time sequences.
Which means as long as you can reframe your problem - with both inputs and associated outputs as a single time series with fixed size time steps - you can give HTM try.
Could you give a brief motivational example for this approach? I think I have a clear picture of the training stage but I’m not sure I fully understand how to set up inference.
The motivation is TM (Temporal Memory) operates with a single time stream, where “inputs” is the sequence of previous time steps, and the “output” is the prediction of one future step.
It does not have separate “input” and “output” streams. So if you want to see “outputs” you have to encode together each
(input(t) and output(t-1)) values in a single SDR and feed them together to TM , which will hopefully predict ((input(t+1) and output(t)) SDR. This predicted SDR you feed into training a linear regressor to see whether it predicts reasonable outputs