When I talked to experts in ‣, I was surprised that most of them said it was just about energy efficiency, and that we could get most of the benefit from them by focusing on that. This is often told as a story by people working on new hardware that uses spiking.
Motor Actions require very precise timing, and that might be a case where spiking matters a lot. People working in Robotics have told me that very short feedback loops are required, and this might enable that.
In Backpropagation, there are two passes, forward and backwards. In general, it’s helpful in algorithm design to be able to have multiple passes. Another way of looking at backpropagation is that two types of information are being passed around, activity and error. It’s generally useful to have multiple channels like that. The Boltzmann Machine model has two passes, the minus phase and the plus phase.
In ‣’s Axon model, the Theta Rhythm is used to create those two phases. So the precise timing of a spike can mark it as being one type of information or another type. One pass or another pass.
Per ‣, “there are multiple codes, spiking is just a highly flexible way to switch between codes, synapses read different codes over different time scales and for different sense modalities”. See Neural Multiplexing
It might be the case that spiking networks implement Variational Inference inference, because at any given time, most neurons aren’t active. I guess it’s kind of like node dropout. But it’s not exactly like dropout, because node activity is correlated.
See ‣