This company has been previously chopped up by
@Diogenese and myself but even this article at the end confirms it’s not coming to a shop near you this decade:
“There are big challenges there,” said
Tony Kenyon, a nanotechnology researcher at University College London. “But we still might want to move in that direction, because … chances are that you will have greater energy efficiency if you’re using very sparse spikes.” To run algorithms that spike on the current NeuRRAM chip would likely require a totally different architecture, though, Kenyon noted.
For now, the energy efficiency the team accomplished while running large AI algorithms on the NeuRRAM chip has created new hope that memory technologies may represent the future of computing with AI. Maybe one day we’ll even be able to match the human brain’s 86 billion neurons and the trillions of synapses that connect them without running out of power”
My opinion only DYOR
FF
AKIDA BALLISTA
I believe we will see, and probably quite soon also, products that utilize NVM (and specifically ReRAM) to "look" like they are in competition with Akida—at the very low end of the market, single purpose, "looks" intelligent kind of thing.
Process In Memory (PIM) is real and exists today. NVM can already be used to simulate a SNN, and ReRAM brings robustness, low power consumption, speed, and appropriately sized technology to the table.
Simulating neurons in NVM has been shown to work (Weebit have demonstrated this). These solutions are fixed in what they can do, but, as with the Weebit demo, single purpose may be all that is needed, as in recognizing hand written, single-digit numerals.
Yes they are programmable, and hence can be adapted to changing situations/conditions (with a new download), but still they are fixed in what they do, and once programmed, are completely autonomous and disconnected from the cloud. Typically, they are single purpose, but for many edge-based sensors, that may be sufficient.
NVM can persistenty store the weights, and change them as a result of prior on-chip calculations (this is a form of learning), PIM can be used to form paths and logic to simulate a special purpose solution that for all intents and purposes is a neuromorphic solution in that it works like the human brain. It will be low power and can be a SNN. Now that WILL find uses. Especially in simple, single-purpose and very cheap products. I even expect that its simplicity may encourage developers to work with this technology. At least they need to learn minimal new stuff and can apply currently software development processes.
That is THE problem Brainchip needs to solve—helping developers learn that Akida is indispensable to them. Make it so they don't want to look anywhere else for a technology to solve their problems. Make it so developers can use familiar processes to arrive at solutions utilizing Akida. Until Akida is more widely entrenched, this is where the company should be focusing most of its efforts right now. Who knows, they might well be—it's just that they are saying absolutely nothing about this side of the business.
Making Akida more powerful is all well and good, and is necessary to maintain a technological advantage. But getting it into the hands of developers and entrepreneurs, and helping them learn how to use it is the current obstacle. To be blunt, who cares how powerful it is if nobody wants that power, or more correctly, nobody yet knows they want that power.
A ReRAM SNN implemented product will probably catch on, as it will be cheap. Much more money will be spent in development, but this will be compensated for by volume sales of a VERY cheap implementation. It still suffers from a lot of the current day CNN limitations such as:
- dependencies on a database of knowledge created by deep learning.
- It will not do single-shot learning nor will it self-adapt to changing conditions.
- It will not be able to do anything more than the single task it is designed to do.
etc.
Using ReRAM is quite a brilliant workaround for not needing to design and build a special purpose chip (now that IS expensive, and time consuming, and has a dependency on fabs etc). Most of the work will be done by software engineers and the implementation will be on off-the-shelf, cheap chips. I think the world will accept and embrace such simple solutions. People will believe these devices are intelligent and their lives may well be enriched by them.
In the end, that may be all that is needed for many AIoT devices.
I'm not saying Akida doesn't have a place, it's just that Akida is so powerful that many are confused by what it can be used for.
) then it’s a good chance someone in the WSU has had their little mittens on one by now.
