It has been taken for granted that 'the only way to do computation is to manipulate silicon in fiendishly complicated billion-dollar factories and then write code to make it work. At the AI Research Group (AIRG), Bill Aronson, CEO of the London based startup explains that they're working on a 'matrix' to identify a cheap material that can perform computation in a normal environment. Today, Bill and his team are using it to build a security device that is about the size of a US quarter coin and will fit on an Arduino shield. It is impossible to hack, clone or spoof. We see immense potential for protecting IoT devices.
At the Deep Learning Summit in Boston this May 24 - 25, Bill will be presenting his current work and how they've created the Matrix, a world first in 'programmable matter'. As well as their work on the Matrix, AIRG is also building a solid-state reservoir computer to shift AI from software to hardware. Apart from being secure, it is at least hundred times more powerful than current AI solutions. They have developed their own proprietary three-layer convolutional neural network (CNN). While most CNN’s have hundreds of layers, they decided to build a three-layer CNN because it would run faster, require less powerful processors and be easier to train. To demonstrate its capability, they have it running on a Windows laptop identifying objects such as cars, bikes and pedestrians in real-time. As an object recognition system, it can be trained for any purpose. Their customers will decide. AIRG just signed an MOU with a company that wants to use it in the manufacture of nutraceuticals. In the run-up to the event, we caught up with Bill to get some behind the scenes information on his talk and upcoming presentation.
How did you begin your work in AI?
This kind of science is a bit like movie making. It attracts very creative people who come together on projects for a fixed period. Some of our team members have been developing AI for two decades or more. We feel we are building a tunnel from two ends. I’m excited to see what will happen now that the material science folks and the AI software developers have met in the middle.
What are the main challenges in your work? How are you using AI to overcome these?
AI is not just software. The stumbling block has been the lack of progress in material science to advance AI systems. For example, security is a big challenge for computing and IoT devices. By reframing the problem using physics we eliminate hacking, cloning and spoofing. It is commonplace for websites to have a challenge-response pair such as the make of your first car to know that it is you. With our material, each IoT device has a unique set of thousands of challenge/response pairs. Every response is different and learning how one device responds doesn’t reveal how another one would behave. It’s so secure you can even transmit the information in the clear.
How have recent advances in AI helped your research?
Perhaps the non-advances are more interesting. Deep Recurrent Nets can be used for processing time signals but there is a significant investment in training using GPU’s and Tensorflow based machines for time domain signals. This is brute force. At the Intel-sponsored NICE2018 in Oregon we saw a shift. We all think of computing as silicon, binary, digital. So, it was good to see the tide turning with lots of attention now on analogue computing.
How are you using AI for a positive impact?
Our approach is a novel analogue computing system capable of processing both analogue and digital time domain signals almost instantaneously. Energy consumption is an issue we tend to forget. Let it be someone else’s problem. Collectively, it is still our problem because it affects our environment, our planet. We are excited by plans to colonize Mars but that’s a backup plan for a few people if all else fails. That’s why we focus on solutions that can dramatically reduce energy consumption.
What developments of AI are you most excited for, and which other industries do you think will be most impacted?
The impacts of AI are complex and will affect every industry. So, we are excited that people are really thinking about the social impacts of AI. We encourage support for The Future of Life Institute which is taking a lead role.
AI and machine learning raise many ethical concerns such as bias, security & privacy amongst others. What are your opinions on this and what can be done to avoid biased machines?
The key to bias is transparency. Until recently the progress of AI has largely been confined to the world of science. Now science and commercial reality are converging. Thanks to Cambridge Analytica we are becoming aware of how AI can even sway elections. In the field of transport, let’s not forget that two people have died recently. One was a pedestrian struck by an autonomous Uber. The other was the driver of a Tesla that veered into a central reservation. These were not accidents but preventable incidents whose cause is to be determined. Car companies take note: the FAA will not certify CNN’s for helicopter because nobody can explain how they arrive at decisions.
So, companies will need to pay more attention to what insurance companies, regulators and legislators require and not just focus on the technology.