Hi chef,well thank you for your response! I had some similar ideas. what is meant by n-of-m coding? do you mean the tensor? I guess they need at least some additional memory and units (sigmoid & MAC) in each NPU to not fall back to CPU and ensure performance and parallelisation. thus I guess I understood the problem why they have to build a completly new chip. thanks!
N-of-M coding is based on the realization by Simon Thorpe's group that the strongest signals are recognized earlier by the retinal photoreceptors of the eye, and by the pixels of a photodetector, and thus generate spikes earlier than the weaker light signals. This means that most of the visual information is carried in the spikes which arrive earlier, and that the later-arriving spikes can be ignored with little loss of accuracy.
the following slides are from ST's presentation
but I don't have the link.
The 4-out-of-16 is just an example as there can be many more than 16 synaptical connexions to a neuron.
Akida does not do MAC.
Late edition:
This patent application, soon to be granted in US, is useful to understand the Akida NPU:
WO2020092691A1 AN IMPROVED SPIKING NEURAL NETWORK
Last edited: