Inside | Real news, curated by real humans
Inside AI

Inside AI (Mar 10th, 2019)

Happy Sunday and welcome to the weekend edition of InsideAI!  I'm Rob May, CEO of Talla, and host of the AI at Work podcast.  Every Sunday we bring you the most popular articles from our daily newsletters this week, and some unique commentary and perspective on things going on in AI.  If you like what you read, please forward this to a friend to help us grow.

The most popular articles from the weekday newsletters...

Leaders from two of China's largest tech companies proposed ethics rules to guide AI development. Baidu CEO Robin Li Yanhong and Tencent CEO Pony Ma Huateng each presented a proposal at the Two Sessions in Beijing, the annual meeting of legislators and political advisors. Ma's plan called for ethical regulations of emerging technologies including AI, and Li urged the government to consult industry experts, businesses, and the public. China seeks to become the global AI leader by 2030. — SOUTH CHINA MORNING POST

This week, the Defense Advanced Research Projects Agency (DARPA) showcased some of the projects from its AI Next program. AI Next is a $2 billion, multi-year initiative to generate the "third wave" of AI technologies. The agency hosted a colloquium on March 6-7 where scientists and technologists could see some of the latest research concepts, including giving machines common sense, teaching systems to learn more rapidly and with less data, and designing more efficient chips. — TECHNOLOGY REVIEW

Forty percent of European companies claiming to use AI don't actually use AI, according to a report by MMC Ventures, an investment firm based in London. The report's authors reviewed the business activities of 2,830 AI startups across 13 EU countries. The report also found that AI firms, or companies that say they use AI, raised between 15 and 50 percent more capital and had higher valuations than traditional software companies. — ZDNET

Japanese startup Vaak developed AI surveillance that detects potential shoplifters. The software scans security camera footage, looking for fidgeting, restlessness, and other suspicious behavior with the goal of preventing crime. The Vaak system is currently being tested in dozens of stores in Tokyo and the company started selling a commercial version this month. Despite the privacy concerns some people have about surveillance, Vaak founder Ryo Tanaka says the platform could go beyond retail and be used in public spaces to prevent crime or suicide. — BLOOMBERG

CogitAI developed a reinforcement learning platform for companies. The California-based startup was founded by AI experts including University of Texas professor Peter Stone, and the "father of reinforcement learning" Rich Sutton serves as an advisor. Reinforcement learning, the type of machine learning used by DeepMind's AlphaGo to master the game Go, is relatively new and the CogitAI platform incorporates cutting-edge algorithms and the ability to apply learning to new situations. — TECHNOLOGY REVIEW

  • Email gray
  • Permalink gray

-- Commentary --

This week I have a Q&A with Osh Agabi from Koniku.  The company builds chips that merge biology and silicon (and, full disclosure, is one of my investments)

1.  Koniku is unique in that you are trying to use biological neurons in silicon to create intelligence instead of the usual approaches.  What was the genesis for the idea?  How long have you been at this?

This idea has actually been around for a while. Some of the earlier works in this direction were done by Jeremy Pines @ Caltech. Some genuinely excellent work was led by Peter Fromherz (he reviewed my Masters Thesis, with I did at the ETHZ in Zurich & Umeå University in Sweden) at the Max Planck Institute for Biophysical Chemistry in Martinsried. Steve Potter at Georgia Tech spent many years building the community. Some excellent characterization work was done by Daniel Wagenaar who was with Steve Potter. 

I did some earlier work in the group of Prof. Henry Baltes at the ETH in Zurich. In grew neurons and did the post-processing of a CMOS chip which interfaced with neurons on a chip. I also introduced some new ideas, which were originally developed by Wolfgang Maass at the TU Graz in Austria. This is a summary, as many more people have helped develop this field. Some genuinely excellent work I have left out - I apologize. This field has a very rich and deep scientific history.

 I have been at this since Q4 2003. I have wanted to start building this technology at startup pace since roughly 2004/2005, but I did feel a lot of the technologies required to really push this into the market and begin to make an impact were not there. In 2014/2015, the landscape changed and I realized the time had come. I formally founded the company in California, August 2015. 

2.  This raises a different set of challenges than other types of hardware, particularly, how do you keep the neurons alive on the chip, and how do you integrate them with silicon from a signal processing perspective?

I couldn't have put it better. Yes. one of those key challenges is how to perfuse neurons on a chip. We have been keeping neurons alive in the lab for decades but within an incubator. So what are the engineering principles of an incubator? There are pH control, temperature control, and humidity control.  Can we shrink this incubator to the size of a smartphone? 

One of the key insights we developed at Koniku is the realization that we are building *a brain* not *the brain*. The Pharma industry needs a physiologically realistic model hence the process they use in that industry. We don't need a physiologically realistic model, we just need to build a "brain" to perform a function. 

We can integrate with silicon by two modes, either reading the optical signals from the cells wit ca imaging or reading the electrical signals with metal electrodes. We have both technologies. 

3.  Startups are difficult because you have to find an initial application.  What are the first applications for a Koniku chip?

Yes. Startups are really hard. This is by far the hardest thing I have ever done in my life. But also one of the most consequential for us - I believe. I also believe autonomous intelligent systems with human quality cognition of the future will be bio-hybrid or cyborg-like - general intelligence if you must. However, I think this is at least 10 years out. Can we go faster? I ask myself that question every day. 

But I concede that the quickest way to get to that destination or impact requires us to plug into the world in a way we can make a meaningful impact. I literally asked myself this question, "what can we do right now that improves people lives in no small measure, demonstrates the power of our platform or superiority to silicon and we have an unfair market advantage?" Turns out, G-coupled protein receptors (olfactory receptors) are really good at detecting volatile organic compounds. That is, biology is really good at smelling. A paper came out in Science in 2017, if I recall correctly, which estimated that humans can discriminate 1 trillion different smells. Compared with 10,000 tones and 7.5 million colors - that is the audiovisual industry, which combined is a 2 trillion dollar industry. 

Here we found our first applications. An application where we can comprehensively outperform silicon machines. For example, we have a device today which can detect explosives in a non-contact mode, which is the footprint of a smartphone.  See it in action here: Current machines can be as big as a small microwave oven and they can detect perhaps 9 - 15 compounds in contact mode. To add more capability to our device, the size or power consumption does not change. 

4.  Do you think neurons fundamentally process things differently than we are ever capable of in pure silicon, or do you just believe this is a faster path to intelligent chips?

Yes. Neurons process signals differently. Also, a neuron does not do a single computation. There is a common misconception that a neuron is equivalent to a transistor. No! A neuron is more equivalent to an FPGA chip...and I don't believe its a faster path, ...going out on a limb here... it's probably the only path. 

Some intelligence is possible with silicon as we can clearly see. I suspect intelligence built with structured data, in a virtual domain will have a hard time competing with intelligence built on sparse data in the real world with massive degrees of freedom. The questions will be settled within a decade I suspect. 

5.  How does using neurons affect constraints you often find in the chip market like power consumption and footprint?  Are you worried that you will have to ignore some applications because you can't scale down neurons as you can CMOS?

The neurons are already scaled down. Take for example the giant squid axon. You could call it an earlier version of the axon or an experimental version...If you realize that the neuron is equivalent to an FPGA class device, then its really small for its size. 

6.  What's the most surprising thing that has come from your work on neuron-silicon integration?

How synthetic you can get with the microenvironment neurons live in. How much modifications you can actually make with this piece of wetware, that has surprised all of us. 

7.  Do you still "program" a Koniku chip in the traditional sense?  Have you, or will you, write software to integrate with existing industry toolsets?

We want to build the Konikore into autonomous systems. The immediate one that comes to mind is robotics for example or mobile autonomous systems. One of the key underlying principles of learning in biological learning is the so-called spike time dependent plasticity. It basically means, if you stimulate two neurons which are connected together within a time window, you can modify the weights or membrane potential between the neurons. The signals or information that are encoded in these stimulus signals can be environmental cues for example. Rewards systems can be another set of signals or even chemical signals delivered within a microfluidics framework. 

The devices which leave our assembly line have a basic set of pre-programmed behavior. The user will train our devices much the same way they will train a pet for example. 

We are toying with other examples or circuit designs which will be going to academia by the end of next year. We will build a community around stimulus standards and best practices which are familiar to most developers. 


8.  Are there spillover effects from Koniku's research?  For example, will the work you are doing help to improve brain-machine interfaces more generally?

Koniku developments, framework or platform will definitely feed into the brain-machine interfaces. If we the leading company building cyborgs. It falls to reason we have an argument in the brain-machine interfaces market. The way I think about it is as follows, Brian machines interfaces companies are building from the top down. Our strategy is from the ground up. 

9.  What is one problem that you aren't working on directly that you wish someone else would solve to make your life at Koniku easier?

Genetic engineering and bio-foundries. A lot can be done to make cells live longer, sturdier, survive even harsher temperature fluctuations and more. It comes down to how cells and their pathways can be tweaked. I think we can be faster at this. Imagine we build computers, and also have to make all own electrical components. Sure, it makes our company more robust or gives us a more defensible position. But I'd prefer to be faster. 

Biofoundries in the sense of folks that can make large amounts of cells to specs. There are hardly any such companies around. 

10.  What's your big vision for where this could go if you are successful?

Cyborgs, ... and that is just a tip of the iceberg :-)...

  • Email gray
  • Permalink gray

Subscribe to Inside AI