There are some issues around M3 that concern me, but which I suspect will cause derision from the assembled masses. For that reason I have been reluctant to write even though several questions have been rattling around my brain for some time.
Background: Brainchip is a hardware implementation of artificial neural networks (ANN). ANN has been well understood and used for 2-3 decades, including the spiking type for at least the last decade.
Proposition: Brainchip does not facilitate any ANN applications that were not previously possible in software implementations. It is not a revolutionary technology. I grant that it does bring speed in buckets, but from an application development point of view, that is all it brings.
Hence my confusion over the proposal that the M3 API is one that operates over the internet. I would normally expect a new development kit to come as a preconfigured board that slots into an existing hardware bus so that the API can exploit the full power of the new chip locally. Why, if speed is your one great advantage, would you dilute that by sharing with multiple users over the internet.
Firstly, I think, this underlines for me the fact that there is no production chip available. The first Brainchip has not been fabricated. Indeed, will the API running over the internet be running on brainchip hardware, or a software simulation thereof.
Next I query if Brainchip can ever be a generalised processor in the way a CPU is. This is because ANN are not generalised but must be configured in terms of number of inputs, outputs, layers and connections optimised for each learning task. I have seen no indication that the Brainchip ANN can or is self configured, and until it is configured no learning can take place.
From this I suspect each chip must be customised for its target application. This would also explain why the business model is to licence a technology that allows each customer to build a customised chip for their own aplication, rather that do an intel, and sell a single generalised CPU.
Now if this implication is correct it has profound effects on the market, and the time to market. Firstly each application would need its own configuration to be established, and a brainchip (ANN) that can learn one task is probably not suited to other tasks. Secondly each chip (ANN) will need to be developed and tested individually which implies an extended time to market for each application.
I am looking forward to seeing the API, and playing with it, as it will give me, for the first time, a real insight into what the technology can do and whether it offers more than existing ANN other than speed.
Disclosure: I have previously held this stock and continue to watch closely as I have had an interest in AI for over 30 years. I would love to see this fulfill its promise - but then I've seen such promises too many times before from too many different AI technologies (including ANN).
Expand
(20min delay)