Thanks homeales. You're correct that I did not adequately account for JAST. I did not take a close look at JAST and how it would impact the design, and had assumed that Akida was continuing the architecture that was previously published in the patents and utilized in the prototypes. Assuming Akida is adopting JAST and the
EP3324343A1 patent is a fair description of the JAST design, this is a significant a departure from the architecture described in the published patents. The synapse goes from being a counter that increments on spikes (and decrements over time) according to configurable parameters, to a binary value that either persists on a match or is randomly set/unset on a mismatch. Perhaps the only congruence between the two architectures is the soma and neuron portions, now working on a count of bits rather than on a sum of synapse values.
Can anybody provide further corroboration on a development path where Akida uses an architecture based on the JAST design instead of the architecture used in the milestones according to the published patents? I had previously gotten the impression that there was more continuity between previous work and Akida, but now I'm doubting that.
How does this impact the concerns I raised? For one, it means that the milestone systems are not very informative of the resource requirements for Akida. That said, it does not make it feasible that there are actually 1.2 million neurons and 10 billion synapses in a $10 - $20 chip, so there still is some unexplained scaling factor being applied to get to those numbers. However, the scaling factor might be quite a bit smaller than I had previously anticipated. I could see that the adjustment to the synapse binary status could be as simple as 4 gates (two ANDs, an XOR, and an OR), which requires about 15 transistors, not counting the logic to support the randomized input for the set/unset step. 15 x 10 billion = 150 billion transistors just for the synapses, so that's definitely not a literal synapse count.
Thanks again for influencing me to dig a little deeper on this aspect of the technology. I would appreciate input anybody can offer on statements or other indicators from the company that would support or undercut the understanding that the SNN design migrated from the old architecture (see published patents) to the JAST architecture for Akida.
(20min delay)
