The First Real AI Neural Net?

[Simo559]

So theres this new emerging company coming out from Australia that has According to what's being said they have created the worlds first AI neural net on a Physical chip and are now selling development kits.

 

Quotes

We are revolutionizing Edge AI applications with our event domain neural processor and comprehensive development environment. Akida™ hardware and software products are the result of over a decade of fundamental R&D by BrainChip engineers and data scientists. Akida AI IP, AI SoC and development environment help customers create ultra-low power solutions with the ability to incrementally learn on-chip without the need to retrain in the cloud.

The BrainChip purpose is to deliver technology that has the capability to improve the human condition through our brain-inspired Neuromorphic AI architecture. We believe that the path to advancing Artificial Intelligence is not by building ever-faster computers, but by enabling specialized chips that can learn like the brain. Our entire team is focused on unlocking the future of AI, with sustainability and performance at the heart of our engineering practices. This is our mission.

 

 

My thoughts

Although i am very excited to see a real working AI that anyone could purchase and program im not certain on how legitimate their claims are, their stock hasn't changed much in the last few months despite signing deals with the megachips corporation in japan and although as of this posting you can go to their website and buy a development kit from them i still remain sceptical if this is really possible technology or if its just snake oil. (really hoping its real).

 

Sources

https://brainchipinc.com/

https://brainchipinc.com/technology/

https://shop.brainchipinc.com/products/akida-enablement-platform-raspberry-pi?pr_prod_strat=copurchase&pr_rec_pid=6995618922677&pr_ref_pid=6995615809717&pr_seq=uniform

https://shop.brainchipinc.com/products/akida-enablement-platform-shuttle-pc?pr_prod_strat=copurchase&pr_rec_pid=6995615809717&pr_ref_pid=6995618922677&pr_seq=uniform

[Quackers101]

sounds like a lot of marketing. not sure how useful this one can be. maybe if they can show something, and not more pigs, elon musk.

 

what sounds more fun is time crystals, with the google quantum machine.

[Salv8 (sam)]

i'll tell ya right now all that device is gonna do is drink beer, shout at the tv while the footy is on and someone has done anything, and have a whinge that the council didn't do the bloody bins this week.

[Doobeedoo]

Make me a replicant already.

[RejZoR]

Only real "Ai" was the one where they actually emulated how neurons behave. Iirc the machine was on intelligence level of a mice at the time.

 

And main reason why I always complain about all the Ai bullshit is that it's all still static algorithmic behavior which is why I can't ever call it "Ai" but just good old algorithms with fancy name smacked on them. Real Ai would ALWAYS store results and always have some benefit from it at a later point if task is similar or same. None of today's neural chips do that. They may perform clever tasks, but the moment they are done with particular task, puff, all the "learning" is gone.

 

Technically, if Ai was done right and neural chips could store their data and reuse it at later time for next tasks, it should technically become faster and faster every time. Today it only gets faster because they give neural engines more execution units that chew through tasks faster than old one. There is no learning or optimization involved on same version of the chip. Because none use store and recall functions in any capacity.

 

Soooo, unless this thing can do exactly that, it's no different than all the neural engine chips we have already.

[BuckGup]

1 hour ago, RejZoR said:

Only real "Ai" was the one where they actually emulated how neurons behave. Iirc the machine was on intelligence level of a mice at the time.

 

And main reason why I always complain about all the Ai bullshit is that it's all still static algorithmic behavior which is why I can't ever call it "Ai" but just good old algorithms with fancy name smacked on them. Real Ai would ALWAYS store results and always have some benefit from it at a later point if task is similar or same. None of today's neural chips do that. They may perform clever tasks, but the moment they are done with particular task, puff, all the "learning" is gone.

 

Technically, if Ai was done right and neural chips could store their data and reuse it at later time for next tasks, it should technically become faster and faster every time. Today it only gets faster because they give neural engines more execution units that chew through tasks faster than old one. There is no learning or optimization involved on same version of the chip. Because none use store and recall functions in any capacity.

 

Soooo, unless this thing can do exactly that, it's no different than all the neural engine chips we have already.

That's actually not true anymore. At Tesla's latest tech conference they shared how the AI model for self driving works and it behaves like a brain. When they want to make it learn something new like a type of sign it doesn't know they'll train a small model first and then the larger model ingests it to learn that skill without the training needed. I'm not sure if it can use it's existing knowledge to make predictive decisions on things it has never seen but AI exists that can do this. So it's only a matter of time before we can incorporate it all in one. 

[MageTank]

11 hours ago, Salv8 (sam) said:

i'll tell ya right now all that device is gonna do is drink beer, shout at the tv while the footy is on and someone has done anything, and have a whinge that the council didn't do the bloody bins this week.

Imagine trying to develop with Australian AI. When I want it to learn "No", do I have to say "Yeah Nah"? Likewise, do I have to use "Nah Yeah" for yes? This won't be easy, lol.

[RejZoR]

1 hour ago, BuckGup said:

That's actually not true anymore. At Tesla's latest tech conference they shared how the AI model for self driving works and it behaves like a brain. When they want to make it learn something new like a type of sign it doesn't know they'll train a small model first and then the larger model ingests it to learn that skill without the training needed. I'm not sure if it can use it's existing knowledge to make predictive decisions on things it has never seen but AI exists that can do this. So it's only a matter of time before we can incorporate it all in one. 

It's not Ai if you're telling it what to do. And that has been my constant gripe with "Ai" being thrown at literally everything. It's not intelligence, it's just bunch of IF statements that "Ai" goes through. If there is something outside of that it's as clueless as a brick.

[BuckGup]

51 minutes ago, RejZoR said:

It's not Ai if you're telling it what to do. And that has been my constant gripe with "Ai" being thrown at literally everything. It's not intelligence, it's just bunch of IF statements that "Ai" goes through. If there is something outside of that it's as clueless as a brick.

I'm confused what you mean then because a model you train isn't a bunch of conditional statements and self learning models do exist. I can train a model to identify a bird so then if it encounters a bird it has never seen before it still has a high confidence it's a bird based upon its current understanding. This means to get better accuracies at identifying birds you need to give it more varying data like different species of birds, color, and background settings so it can take these factors into account. It's the same way human models work.  

 

Written out it appears to be conditional statements like identifying a bird

 

I see object outside move on it's own-> must be living

Object is small and doesn't look human -> must be an animal

I see animal has two legs and beak -> must be a chicken

Animal begins to fly -> must be a bird not chicken

 

but in reality you are more so consulting your internal model to fact check yourself or have greater prediction accuracies. So it's not one specific component but the entirety of the model. 

 

It's like the quote, "It's a bird, it's a plane, no it's superman!" or how optical illusions exploit fundamental pieces in your models to trick you.

 

I think it's also why kids are so awed by dinosaurs as they take models like birds only having beaks and showing them animals with beaks that ARE NOT birds that were real

[Salv8 (sam)]

2 hours ago, MageTank said:

Imagine trying to develop with Australian AI. When I want it to learn "No", do I have to say "Yeah Nah"? Likewise, do I have to use "Nah Yeah" for yes? This won't be easy, lol.

thats if you get it to stop talking about scomo and his shit polices.

[RejZoR]

2 hours ago, BuckGup said:

I'm confused what you mean then because a model you train isn't a bunch of conditional statements and self learning models do exist. I can train a model to identify a bird so then if it encounters a bird it has never seen before it still has a high confidence it's a bird based upon its current understanding. This means to get better accuracies at identifying birds you need to give it more varying data like different species of birds, color, and background settings so it can take these factors into account. It's the same way human models work.  

 

Written out it appears to be conditional statements like identifying a bird

 

I see object outside move on it's own-> must be living

Object is small and doesn't look human -> must be an animal

I see animal has two legs and beak -> must be a chicken

Animal begins to fly -> must be a bird not chicken

 

but in reality you are more so consulting your internal model to fact check yourself or have greater prediction accuracies. So it's not one specific component but the entirety of the model. 

 

It's like the quote, "It's a bird, it's a plane, no it's superman!" or how optical illusions exploit fundamental pieces in your models to trick you.

 

I think it's also why kids are so awed by dinosaurs as they take models like birds only having beaks and showing them animals with beaks that ARE NOT birds that were real

The model they train. And is then eternally stupid with those baked in parameters. Aka, not really intelligent. I mean, are you eternally stuck with static information about everything or do you learn about things with every second of being alive without the disconnect between doing things and learning?

[butcherano]

29 minutes ago, RejZoR said:

The model they train. And is then eternally stupid with those baked in parameters. Aka, not really intelligent. I mean, are you eternally stuck with static information about everything or do you learn about things with every second of being alive without the disconnect between doing things and learning?

That’s not how Brainchip’s Akida works.  It has the ability for on-chip learning without the need for going back to the cloud or retraining the entire network every time it comes across something new.  This is what distinguishes it from any other product on the market which are essentially just mathematical accelerators.

 

I’d highly recommend viewing the links in the original post and doing some thorough research on Akida.

 

This is probably the best summary video which has some great examples in it as a good starting point…

 

And a good analyst report here…

 

https://static1.squarespace.com/static/5af533a312b13fb602fe7d7b/t/612313b3aae13c19f15e79bb/1629688762738/BrainChip+Holdings+research+initiation+report+-+20+08+2021.pdf

 

 

[Kisai]

22 hours ago, Simo559 said:

 

The BrainChip purpose is to deliver technology that has the capability to improve the human condition through our brain-inspired Neuromorphic AI architecture. We believe that the path to advancing Artificial Intelligence is not by building ever-faster computers, but by enabling specialized chips that can learn like the brain. Our entire team is focused on unlocking the future of AI, with sustainability and performance at the heart of our engineering practices. This is our mission.

 

I sincerely doubt this is a "neural net" a la star trek, but more likely just a dedicated ASIC with something similar to GPT-3 or something in it.

 

Like, don't get me wrong, the entire purpose of AI research is still very wild-wild-west, so "better" hardware that comes out will likely be a specific-purpose. For example the hardware that is designed for AI-conversational and AI-Dungeon would be similar, but OpenCV and Tacotron are pretty far apart from each other and GPT.

 

 

[100001]

1 hour ago, Kisai said:

I sincerely doubt this is a "neural net" a la star trek, but more likely just a dedicated ASIC with something similar to GPT-3 or something in it.

 

Like, don't get me wrong, the entire purpose of AI research is still very wild-wild-west, so "better" hardware that comes out will likely be a specific-purpose. For example the hardware that is designed for AI-conversational and AI-Dungeon would be similar, but OpenCV and Tacotron are pretty far apart from each other and GPT.

 

 

 

It's not a dedicated ASIC. The chip is algorithm/network agnostic. The same chip can run networks that process data from all sensor modalities.

 

its a new type of processing architecture that's design is based on the biological neuron. Each processing node mimics the functionality of a neuron and synapses. The processor is a SoC with a neural fabric consisting of lots of these nodes (neurons) and and ARM m-class processor for managing the neural fabric.

 

For example, BrainChip have a model zoo that provides a bunch of different networks to run on the device out of the box:
https://doc.brainchipinc.com/zoo_performances.html

 

If you wanted to try out the one-shot learning capabilities, I built a basic boilerplate app in python to learn how to use it:

https://github.com/stdp/akida-camera

[igormp]

On 12/28/2021 at 10:44 PM, Simo559 said:

Although i am very excited to see a real working AI that anyone could purchase and program im not certain on how legitimate their claims are, their stock hasn't changed much in the last few months despite signing deals with the megachips corporation in japan and although as of this posting you can go to their website and buy a development kit from them i still remain sceptical if this is really possible technology or if its just snake oil. (really hoping its real).

That's not how it works, this is a product made for devs to build edge devices that can perform inferences. The nice feature of this one is that you can basically do some kind of "online training" without needing to rely on external services.

 

On 12/29/2021 at 12:57 PM, RejZoR said:

Technically, if Ai was done right and neural chips could store their data and reuse it at later time for next tasks, it should technically become faster and faster every time. Today it only gets faster because they give neural engines more execution units that chew through tasks faster than old one. There is no learning or optimization involved on same version of the chip. Because none use store and recall functions in any capacity.

That's called a cache, and it exists. If you mean trying to learn the stuff from current inferences without needing to retrain a new model offline, it also exists already and it's called online learning.

 

On 12/29/2021 at 9:25 PM, Kisai said:

I sincerely doubt this is a "neural net" a la star trek, but more likely just a dedicated ASIC with something similar to GPT-3 or something in it.

 

Like, don't get me wrong, the entire purpose of AI research is still very wild-wild-west, so "better" hardware that comes out will likely be a specific-purpose. For example the hardware that is designed for AI-conversational and AI-Dungeon would be similar, but OpenCV and Tacotron are pretty far apart from each other and GPT.

No, it's more akin to a mix of a tensor multiplication device and an FPGA. Whereas with something like a Coral accelerator you'd have your model stored somewhere, and use the matrix units to perform inferences, with this device you have a configurable mesh (pretty much like an fpga), but that's limited to only performing actions like a single unit in a model.

However, unlike most FPGAs, it's easy to readapt that mesh on-the-fly, apparently.