2022 BRN Discussion, page-5242

@Fact Finder ,

re your ref to:
https://arxiv.org/abs/2106.08921

Cor blimey, g'vnor,
That's a bit above my pay grade at this time of night ...

https://arxiv.org/pdf/2106.08921.pdf

At the bottom of page 2, they do say:
For some applications, Loihi shows a clear advantage in dynamic power usage when compared with conventional hardware like CPU and GPU, though the time required to process each input is often higher on Loihi*.

[*Loihi 2 is 10 times faster than the original.
"Faster circuit speeds. Loihi 2’s asynchronous circuits have been fully redesigned and optimized, improving on Loihi down to the lowest levels of pipeline sequencing. This has provided gains in processing speeds from 2x for simple neuron state updates to 5x for synaptic operations to 10x for spike generation."
https://download.intel.com/newsroom/2021/new-technologies/neuromorphic-computing-loihi-2-brief.pdf ]


Basically, they have compressed the input image by reducing the pixel count which produces an evenly distributed pixel reduction. This is quite different from event-driven sparsity used in Akida.

They do point out some of the problems with Loihi 1.


2.4 Firing Rate Regularization
The firing rates of neurons in an SNN impact the power and performance of the network significantly. If a neuron is spiking slowly, it takes longer to propagate information to the next layer of the network. If a neuron is spiking rapidly, on the other hand, performance improves but we lose the advantages of temporal sparsity. This is particularly problematic on Loihi, where the number of synaptic updates is directly proportional to the number of spikes going into a layer, and synaptic updates have a significant cost for both processing time and power consumption. Furthermore, the fact that Loihi neurons reset to zero after a spike means that given a constant input, a neuron will always fire with a period that is a constant integer number of timesteps. This means that at firing rates closer to the maximum rate of the inverse of the timestep 1 /∆t , there is less resolution (for example, if ∆t = 1 ms, then a neuron can either fire at 500 Hz or 1000 Hz, but not at any rate in between). Based on these criteria, we have found that maximal firing rates in the range of 50-200 Hz work well to balance accuracy with energy efficiency and throughput.
...
Instead of using a fixed firing rate target for all neurons on all examples, we make two key modifications: we allow a range of firing rates, and we regularize a rank-based statistic computed across a neuron’s firing rates on multiple examples

Something that puzzles me is that, in a recent interview, the interviewer asked Peter about comparing Loihi and Akida, saying: "They are both digital", and Peter, in passing, agreed and went on to point out Akida's advantages. Now this paper describes Loihi's neuron operation in terms of summing voltages and currents, which I would classify as analog.

On page 10 the paper sets out several equations for the firing of a neuron:



On the other hand, in some Intel patents, the neurons are described in digital terms. So, because of the differing information about Loihi neurons, my question is has anyone here reverse engineered Loihi 2?