ICWA Presentation Notes

As requested by a few people, I've created this thread as an easy reference point for my personal notes on the ICWA BrainChip presentation.

---

For those of you that are interested, here are my notes on last night's presentation. It's not a full account by any means, but I wrote down as much as I could at the time. If you feel that my notes are not adequate in any way, feel free to press the ignore button on my profile.

Overview of Peter's speech, Q&A is below.

His initial research in 2005 led him to come to the idea of a hardware-only solution.
In 2007 he had his 'eureka' moment, with a chip loaded with about 20 neurons. This used dynamic synapses, in a simplified neural network design (Not biologically accurate). Worked with spikes in, spikes out.
In 2008, he filed his patent and had a working design based on sensory input. The STDP learning has been improved remarkably since that initial patent.
Since 2008, he has used larger FPGA devices (upto 1 million gates) to develop BC further.
2010 - further development on artificial neurons. These are not like existing software neural networks, in that they learn from sensory input. He has developed temporal, spatial and motor neurons - not all neurons are created equal. There may be more that BC can mimic but he only named these 3. BrainChip's neurons are configurable to act as a number of types of neurons.
1 BC neuron is as powerful/useful as 4997 sigmoid neurons (the type compared in the video).
2013 - BC learned to recognise white-noise. When BC was presented to neuroscientists it was described by them as 5-10 years ahead of anything else available at the time. This comparison is still valid today, if not further ahead. In conversation, Peter remarked that IBM's TrueNorth is only just now using the basic neurons that he was using in 2006.
In the demo to a group of neuro scientists, Peter showed them the output signals from BC. They all were unimpressed, saying that they've all seen organic neuron signals before. When he explained that they were looking at BC output, they were all amazed that the output was identical in structure to the human neuron.
In 2014 the San Diego/LA office was opened, right in the heart of 'brain valley'. It's there that Peter met Geoff Gershmar(sp??) who is working with 3000 neurons. His work requires 2400 processors to run the 3000 neurons. BC currently has 10,000 on a chip (it takes 0.5 watts to power the 10K neuron brainchip).
2015, the software-car demo is out, and it learns to stay in the middle of the track by using 14 neurons (1 motor, 7 spatial and 6 temporal) - I think these figures were correct, I was too slow taking notes. There is no programming/software involved, the car is driven forward by the single motor neuron, reacting to the 13 temporal and spatial neurons. I confirmed that it was definitely 15 seconds to learn the track. Peter said they had to slow the neurons down for the demo because the animations would have been too quick to see at full speed. The computer could not respond fast enough at full speed so they slowed the BC down.
Spikes are described as events like touches to the skin (in spatial neurons). 1 million neurons will have 100 million associated synapses on the BC. Neurons are connected, so that the output from 1 neuron feeds neighbouring neurons. Spikes generate current, and illicit either an 'exite' reaction or 'inhibit' reaction. If exited, they activate joined-up cells. If inhibited, they disable joined-up cells. The Axon joins neurons together (in the human body, they can be as long as a meter - from your head to your leg).
The BC will consume .5 watt for a 10,000 neuron chip. 88 billion neurons in the human brain. BC can be built to contain upto 260 billion neurons. In the current design, there are 65,000 neurons.
Peter has spoken to 'a number of companies in Silicon Valley'. I wasn't able to get him to discuss investors or partners in the US though.
Since joining ICWA, he has been building a team near San Diego - in conversation he mentioned that a nearby chip maker is closing down and they're recruiting directly from there. One employee was responsible for the PCIE that connects BC to a computer, didn't catch his name - Ish Sherish maybe?? Not sure.
'The science is done' - Peter indicated that it's now time to start commercialising.
Why use AZK to backdoor? Access to capital to fund development, without losing control of ownership. It's the first step towards a US based listing. I asked why not just list in the US now, and Andrew Jones replied that the regulations are much stricter in other markets, and the ASX have recently loosened their regulations to accommodate smaller listings. Also, listing on the ASX through AZK legitimises the brand to global names. BC would not qualify for listing on their own on the ASX. Peter also said that he's Australian, and chose the ASX based on that. He met Andrew Jones through the ICWA so I think this means the ASX would have been the first and easiest option since they're all Perth/Aus based. I think he credited ICWA with most of the recent progress in developing the business side, but can't be sure cos it's not in my notes.
'Building products that are beneficial to people...in a different way' - nice little finishing quote that suggests Peter is dedicated to his invention.

- Questions asked -
Are they using the ARM model? He answered this twice - YES.
How many neurons would it take to do voice-recognition in a phone? - 2800 BC neurons.
Why not just go direct to Silicon Valley - Wants to keep control of IP and grow the business so that BC is available to as many partners as possible. Going to Silicon Valley for funding would mean giving up equity for funds to develop the business from here.
Software API and/or Hardware API? They're calling it the BDK - Yes, available to anyone who wants one. The hardware version will be an ASIC with 10,000 neurons on it, at a low cost (less than $100 maybe).
What's next for BC development? 'Making it faster, we have all the precision we need' - The race car demo neurons were running at 8Mhz, the next step is to run at 100MHz.
Also, there are 16 patents currently being developed by BrainChip. One in particular is related to the export/duplication of BC learning from chip to chip.
How do know the chip can be produced by manufacturers? - They use FPGA's (costing $80K each) that virtualise chip designs into working models. This proves the architecture is physically possible and the design can be printed onto silicon. This uses existing techniques and can be done today.
What other experiments are you keen on? Peter wants to buy an iCub (http://en.wikipedia.org/wiki/ICub) that is currently powered by 60 blade servers. He wants to use a BC to run it.
Licensing plans - Get the BDK out into the world, and then licence any application that commercialises. Charge a percentage per chip - 5c, 3c per chip.

- Questions I wasn't able to get to -
Where did the 2013 funding come from? Sorry, didn't get this one out to Peter. I did spend about an hour talking to he and Andrew and felt that someone else should probably have a turn.
Current Partnerships/companies involved? Wasn't able to ask directly but he did say that a Belgian or German (can't remember) cochlea implant company have been working with them. Coincidentally, the cochlea implant and BC use the same communication method, so are plug-and-playable. Only quote I could record was 'a number of companies in Silicon Valley'.
Is CR available to shareholders - I didn't get a chance to talk to Niall about this but I did overhear him say that they're not intending to do another CR, they have enough money to continue developing the company, and that the current CR was already oversubscribed.

If you're still reading this, I applaud you!