Hello all, newbie here. This isn’t an HTM specific question, more a broad implementation design question relevant to any/all neuromorphic designs. The engineering side of me sees synaptic connections as the major roadblock to any large scale hardware implementation (which is stating the obvious in this forum). Currently this manifests as increased memory utilization (to store data about all synapses). Is there any reason we couldn’t use RF to tackle this?
I found very few papers on this topic from 2006 and 2009 but they’re behind a paywall. I envision each neuron designed with a very small, very low power transceiver. (Of course everything in my head may be far more challenging in reality.) I did some research on LTE forever ago so I’m going to steal concepts from there. If you multiplex in time and frequency, considering the short distances involved, you could potentially get millions of unique blocks in a 10ms, 15MHz window. Every neuron would need to know when and where to look, so there would likely need to be a fairly complex scheduler. The more complex part would be that an activated neuron would need to generate a pulse to every neuron with a synaptic connection (meaning multiple time/freq blocks). The potentially cool thing about this is signal strength. Signals from closer neurons would arrive stronger, and multiple signals from distant neurons could combine to overcome an activation threshold. Another way to handle permanence values could be adaptive filters, where less and less of a specific freq would be canceled out, as in increasing permanence.
Anyway, just something I’ve been mulling over. I know this isn’t fully solving memory reliance. Thoughts?