Well, if the computations required by such a spiking network are inconvenient on a CPU or GPU, then I understand why you want a specialized chip, but quick googling doesn't really enlighten my as to what exactly makes CPUs/GPUs so bad at these networks
I mean, we have massively parallel GPUs because the CPUs can't handle the large data throughput and parallelism needed for graphics processing. What's it about these networks that neither the CPU nor the GPU are suited to them?
"When a neuron’s activation exceeds some threshold level, it generates a spike message that is routed to a set of fanout compartments contained in some number of destination cores. "
yeah, definitely research test chips atm
I could see how if all you wanted to do was simulate a SNN, then a hardware designed specifically for that could be more efficient than simulating one using software...
you'd have to simulate things like spike trains using arrays or something when you'd have a single neuron that can handle all of that in one of these chips...
Huh, this article from 2017 seems to say the main advantage of neuromorphic chips over simulating on a traditional chip wouldn't be speed but power-usage
yes. they are charge integrators with a threshold. This means that if I have a small capacitor and a small current, I can have very long time constants with very low currents.
For instance, I can do key word recognition in 5nW (measured) with neuromophic approaches, or I can do it in 10mW in digital. It's also very application specific.
I am the "honorable mention" for doing most of the IC layout.
In analog, you get natural log mathematics for "free". Most of the natural processing algorithms require this mathematics. This is also very difficult for digital computers to do.
As "Moore's Law" (which isn't a law) has been in the cross hairs of the realities of physics, neuromorphic approaches have become more popular. Carver Mead (who coined the term Moore's Law, who also had a good chunk of Intel stock) wrote a book in 1989, "Analog VLSI and Neural Systems". That was the starter and is very approachable.
@bdegnan So how programmable are these analog networks? Can you have a "generic" network or do you need to design the circuity with a specific application in mind?
The analog approach works pretty well: youtube.com/watch?v=XVR5wEYkEGk If you can get to Tobi Delbruck's lab at ETH Zurich, you can see some silicon neuron-based cortex in action. I like the pencil balancing myself.
Halser's paper shows off some of these circuit techniques.
The floating gate (think FLASH but more bits) will start to generate hot-electron due to impact ionization when the drain moves out of a programmed area. This will change the weight of the device. This means that if the spike is infrequent, it'll have more weight (generally)
Depends. Delbruck uses XFAB in Germany (europe somewhere), Hasler uses TSMC out of Taiwan. I used IBM microsemi in Vermont/New York USA. I haven't made any ICs since IBM became Global Foundries in the neuromorphic space.
I made a single silicon neuron as a proof on concept on a 14nm SOI process from IBM. It worked, but you cannot publish over a single neuron. Patterned metals and other process quirks were challenging.
@bdegnan googling "darpa attentive binocular" does not seem to bring up anything but see nearby prjs eg "darpa + spiking neural networks" cacm.acm.org/news/…
@bdegnan yes maybe some of the idea/ enthusiasm is that possibly one SNN unit could require less electronics/ space/ power than one ANN (roughly speaking). and maybe there is more emphasis on analog (SNNs) vs digital (ANNs)...
The Cognitive Technology Threat Warning System, otherwise known as (CT2WS), is a brain-computer interface designed to analyze sensory-data and then alert foot-soldiers to any possible threats, passive or direct. CT2WS is part of U.S. Department of Defense's effort to produce an efficient and working Network-centric infantryman.
== Project ==
=== Proposal ===
Proposed in early 2007, DARPA came to believe that a visual warning system could be produced and developed via an integration of technology and artificial intelligence. By combining discoveries in flat-field, wide-angle optics, larg...
How that actually was used, you put helmets on people and they watch movies at 24 frames/sec. You insert 12 images per second and then you read out the spikes when they see a "tank", "plane", etc.
All I can say is that the FPGA version of the neural networks always worked differently from the analog version with silicon neurons. They never came to the same conclusions. This does not mean they are incorrect, but the analog ones you can hook into a rat brain, the digital ones never really worked. That was outside of my direct work as it had to be done in France, but they used ICs out of the lab.
If you look for me on google scholar, you'll see that I do unpopular, weird things that don't have much value, but if you look carefully, you'll see that I'm well funded for the ridiculous.
Yeah, I did temperature stability for these circuits.
no, You want Stephen Brinks Ph.D dissertation that I liked to above.
I'm just a semiconductor physicist. I am a means to the ends for the neuromorphic people.
Yeah, it's terrible. I got in a fight with my advisor over a company that I started. There are two types of dissertations: excellent and completed. Mine is completed.
much worse. My boss didn't sign my dissertation for two years after I defended because they needed someone to run the lab while they went to Thailand for some "updates". me: "what do I call you?" them: "Jennifer, and I want babies".
I got beaten up the Hell's Angel's. Got shit drunk with Lil' John and vomited in someone's Ferrari. Had the YAKUZA over from Japan for a party. I'm bad news.
this is all very cool and reminds me of this older chat session AMA idea we did a few yrs ago. think youd be perfect if you have some spare time sometime. for some reason they tended to work better in summers... physics.meta.stackexchange.com/questions/7783/…