We currently use the same SynapseConfig for all synapses in a Connection. This means they all have to have the same values for many parameters, including the weight exponent. Allowing different weights to use different SynapseConfig objects would mean that we can e.g. have a different weight exponent for large weights versus small weights, allowing for more bits of precision for smaller weights. This should result in more accurate networks, and potentially even allow us to use fewer bits to store each weight.
This will make things more complicated, since we'll have to keep Loihi constraints in mind. Not only are we constrained on the number of SynapseConfig objects per core, but there might also be constraints on what compression types this could work with (and possibly other constraints, too).
We currently use the same SynapseConfig for all synapses in a
Connection. This means they all have to have the same values for many parameters, including the weight exponent. Allowing different weights to use different SynapseConfig objects would mean that we can e.g. have a different weight exponent for large weights versus small weights, allowing for more bits of precision for smaller weights. This should result in more accurate networks, and potentially even allow us to use fewer bits to store each weight.This will make things more complicated, since we'll have to keep Loihi constraints in mind. Not only are we constrained on the number of SynapseConfig objects per core, but there might also be constraints on what compression types this could work with (and possibly other constraints, too).