When we read the transcript of Anil Mankar’s presentation to the Ai Field Day he is at pains to point out that the number of frames per second and the number of pixels making up the image is irrelevant as AKIDA is not processing frames but events in those frames instantaneously. AKIDA targets two application areas. One is the sort of conventional convolutional neural networks that are being applied to literally thousands of different benchmarking problems -
https://paperswithcode.com/datasets. For that, BrainChip invested a lot of effort to develop software that takes almost any CNN and adapts it to run on the AKIDA chip, replacing expensive floating point arithmetic with cheaper 4-bit, 2-bit and even 1-bit calculations. It's the CNN2SNN toolkit
https://doc.brainchipinc.com/api_reference/cnn2snn_apis.html. They provide detailed information for several well known models
https://doc.brainchipinc.com/api_reference/akida_models_apis.html#model-zoo. This all makes good sense, but there will definitely be limits to what can be done on a single chip.
The second area is true spike-based processing that could be used with things like Dynamic Vision Sensors that send asynchronous spikes. This is indeed revolutionary and NVIDIA offer nothing for that sort of problem. Loihi and TrueNorth could, in principle, handle this sort of data, but everyone agrees that they are not a commercial proposition. Unfortunately, Brainchip's documentation for these aspects of the AKIDA chip is even more limited, and again, it would appear that no benchmarking data is available. In one of his presentations, Peter Van Der Made said that a single neuron can have up to 160,000 spking inputs, and in principle, they can use the JAST learning algorithms (that they licensed in 2017) to learn repeating patterns. It would be great to have more details on this.
For both classes of problem, it would be very useful to provide some benchmarking figures. Anil's statement that the number of frames per second and the number of pixels making up the image is irrelevant is strange. Much of BrainChip's marketing has been targeting precisely the sorts of CNN applications where such comparison data is vital. I really think that AKIDA would beat the competition on this - certainly for power consumption. But they need to provide some numbers.
(20min delay)