Among the problems you get with the neuromorophic chips is that they are choosing which hypothetical neural algorithm the chip should use and which silicon implementation should be made of that algorithm.<p>Current AI is continuously modifying it's algorithms. Personally, I'd want to see parallel chips with a programming interface at least as "general purpose" as the GPU.<p>One might even boldly say something like "the only way this deep learning explosion happened is through the consumer market making powerful parallel chips ubiquitous". Or more cynically, "if chip maker had been looking at only AI applications, they would have strangled the deep learning explosion by their tendency to demand $10K+ per installation since they look at the well-healed corporation that easily afford these and thus ignore the graduate students who make the conceptual progress in their basements." Even Nvidia is eager to divide their offerings between $500 game boards and $5000+ deep learning boards. One assumes a neuromorophic chip would be way up this "high end".
Transmitting analog voltages in neuromorphic circuits reminds me of a cool experiment I heard about a while ago. A team used an evolutionary algorithm to adapt an FPGA to recognize sounds. When they got the behavior they wanted, they tried copying the circuit to another FPGA and found that it didn't work because the algorithm had evolved dependencies on specific quirks in the original physical FPGA. I can't figure out how to link to the PDF directly, but the original research is probably "An Evolved Circuit, Intrinsic in Silicon, Entwined With Physics" by Adrian Thompson, which tells the story far better than I have.<p>It's neat to think that our own brains might have adapted wiring that works specifically for the actual molecules that make up our bodies. It's uniqueness on a different level.
On this topic, more specifically of whole-brain emulation, a very interesting report I found a few months ago: "Whole Brain Emulation: A Roadmap". It's one of a very rare breed: an evaluation of a <i>future</i> technology that <i>does not</i> enter into wild speculation or wishful thinking, but remains critical and objective as far as possible.<p>[1]: <a href="https://www.fhi.ox.ac.uk/brain-emulation-roadmap-report.pdf" rel="nofollow">https://www.fhi.ox.ac.uk/brain-emulation-roadmap-report.pdf</a>
Another interesting kind of hardware they have at INI (Giacomo Indiveri's institute) are the dynamic vision sensors (initially developed by Lichtsteiner, Posch, Delbruck, Berner, Kramer).<p>Those are cameras modeled after the human retina, and they overcome the processing and frame rate bottleneck of traditional cameras by using an event based architecture (event stream of pixel brightness changes -> no data if nothing changes). This allows microsecond latency in machine vision systems, without the high data rates you get with frame based cameras.<p>* Overview: <a href="http://siliconretina.ini.uzh.ch/wiki/index.php" rel="nofollow">http://siliconretina.ini.uzh.ch/wiki/index.php</a><p>* Company selling them: <a href="https://inivation.com/dvs/" rel="nofollow">https://inivation.com/dvs/</a><p>* Another company selling them split off by one of the PHDs: <a href="https://www.prophesee.ai/" rel="nofollow">https://www.prophesee.ai/</a><p>* Demo video: <a href="https://youtu.be/jnzPuDUsP4w?t=14" rel="nofollow">https://youtu.be/jnzPuDUsP4w?t=14</a>
This neuromorphic stuff is super cool. I wish I could just buy some! I'm the author of spikeflow (github.com/colinator/spikeflow). I'd love to support neuromorphic hardware.<p>For the moment, biological neural networks (brains) are, I believe, many orders of magnitude more powerful, and especially more efficient than, current computers. But it's hard to say, because it's so hard to tease out the important parts of what the brain does, because it does so much. There are so many types of neurons, so many types of connectivity, so many interacting neurotransmitters, so many support cells that affect computation... almost like the brain is the result of a billion years of hack-upon-hack evolution...