A few months ago I read an article saying DARPA had issued one of their requests for proposal for a neuromorphic computing architecture. I thought the request was extremely interesting because it was asking for proposals for how to build a system that was capable of Hebbian like learning and ultimately high levels of cognition, could have as many synapses as a human brain, fit in quite a small package, and consume very little power. I.e., the makings of a potentially cheap, portable, low power, AGI.
The proposal wanted an architecture that could promise of scaling of an average density of synapses of something amazingly small, such as, if I remember correctly, one thousand per square micrometer, or roughly 32x32 nanometers to model each synapse and somehow includes its connections. That would enable 100G synapse per cm2, or 100T synapse in roughly 1 foot square, with just one layer. I Googled a little on the subject and found that most of the electronic analogue models of synapses I found on the web had something like five transistors per synapse, so I though to myself maybe DARPA was thinking of memsisters, which I had read could remember at least one variably selectable analogue value at each position in a densly packed cross bar --- and I had heard a speaker from HP say they felt 100B bits per cm memories were possible with their cross bars (in pre-memsistor days) even including addressing and drive electronics and interconnect to the outside world. When I saw the date that the very lengthy response to the request for proposal had to be filed by was only several weeks after the proposal had been announced, I knew that this request --- as Ben has told me many DARPA request for proposals are --- had been designed with the winner of the grant in mind. I guessed it might be HP with memsistors. I was right. The below quote and its link to the article it came from indicate that DARPA has selected a memsistor approach to its neuromorphic computer initiative, and HP is one of the two contractors. ========== from http://www.eetimes.com/news/latest/showArticle.jhtml?articleID=212200673 "Also at the symposium, Snider unveiled a design that used memristors in their analog mode as synapses in a neural computing architecture. Memristor crossbars are the only technology that is dense enough to simulate the human brain, Snider claimed, adding that the HP Labs crossbars are ten times denser than synapses in the human cortex. By stacking crossbars on a CMOS logic chip, variable resistance could mimic the learning functions of synapses in neural networks. "HP Labs and Boston University were recently awarded a contract by the Defense Advanced Research Projects Agency to build the first artificial neural network based on memristors. " ====== As the article indicates, each layer of these cross bars do not require a silicon substrate, so it is possible to stack many such layers; If this technology pans out, it could make AGI's one whole hell of a lot cheaper to build. Ed Porter ------------------------------------------- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=120640061-aded06 Powered by Listbox: http://www.listbox.com
<<attachment: winmail.dat>>
