Imagemap

Dharmendra Modha Leads Creation of Brainlike Computing Architecture

IBM researcher Dharmendra S. Modha has led the development of TrueNorth, a radically new computing architecture that mimics the efficiency and cognitive capabilities of the human brain.

“It doesn’t make sense to take a programming language from the previous era and try to adapt it to a new architecture,” explained Modha. “It’s like a square peg in a round hole. You have to rethink the very notion of what programming means.”

As part of its pioneering work toward developing more brainlike computing, in 2011 IBM researchers had unveiled chips that use a network of “neurosynaptic cores” to mimic the way neurons work. TrueNorth is the scheme they devised to task those chips for specific cognitive tasks like a visual sensor that can process images the way the human brain does.

To do that TrueNorth distributes the processing of information into vast numbers of parallel paths in the manner of the brain’s neurons and synapses. This is a radical departure from the conventional Von Neumann architecture in which information is retrieved from storage and processed in a single sequence, or at most, in a small number of sequences.

But powerful computers using a massive number of neurosynaptic cores remains theoretical at this point. So a key intermediate step toward their development is using a conventional supercomputer to simulate the functioning of a brainlike machine. Modha’s team has developed software that uses a conventional supercomputer to simulate the functioning of a brainlike machine with 100 trillion virtual synapses and 2 billion neurosynaptic cores in a single massive network.

The simulated neurosynaptic computer is made up of many cores, each with its own network of 256 “neurons” that mimic biological neurons. Each one develops its own response times and firing patterns to process input from neighboring neurons, unlike conventional processors that run at a fixed speed.

The cores are programmed with corelets that determine the basic functioning of a network of neurosynaptic cores. Individual corelets can be nested into more complex networks “like Russian dolls”, says Modha. Each of TrueNorth’s 150 corelets are pre-designed for a specific cognitive function like detecting motion or sorting images.

Mimicking the brain would allow computers to solve problems that would require them to learn, cope with ambiguity, infer relationships and make predictions the way intelligent human beings do.

One contemplated use of TrueNorth is to develop systems that can mimic the human visual processing power to create a visual sensor sophisticated enough to provide intelligent input to self-driving vehicles or robots built to rescue people from burning buildings, for example.

Dharmendra Modha founded IBM’s Cognitive Computing group at IBM Research – Almaden. As the principal investigator for DARPA SyNAPSE team globally he leads a global team across neuroscience, nanoscience and supercomputing to build a computing system that emulate the brain’s perception, cognition and decision-making powers.

Modha received a bachelors in computer science and engineering from Indian Institute of Technology Bombay and a PhD in electrical and computer engineering from UC San Diego.

---

Comments

DR. EDWARD SIEGEL · Dec 29, 06:27 PM · #

Not new and not “news”!!! Artificial neural-networks(ANN) patterned on biological neural networks(BNN) artificial-intelligence(ANN) were alive and well long before 1980 when physicist Edward Siegel [consulting with Richard Feynman(Caltech) for ANN AI pioneer Charles Rosen(Machine-Intelligence) & Irwin Wunderman(H.P.) & Vesco Marinov & Adolph Smith(Exxon Enterprises/A.I.) discovered trendy much-hyped “quantum-computing” by two-steps: (1) “EUREKA”: realization that ANNs by-rote on-node switching sigmoid-function 1/[1 + e^(E/T)] ~ 1/[1 + e^(hw/kT)] ~ 1/[+ 1 + e^(E/T)] ~ 1/[ + 1 + e^(hw/kT)] is Fermi-Dirac quantum-statistics 1/[1 + e^ (E/ T)] ~ 1/[1 + e^(hw/kT)] ~ 1/[+ 1 + e^(E/T)] ~ 1/[ + 1 + e^(hw/kT)] = 1/[e^(hw/kT) + 1] dominated by Pauli exclusion-principle forcing non-optimal local-minima(example: periodic-table’s chemical-elements) forcing slow memory-costly computational-complexity Bolltzmann-machine plus simulated-annealing, but permitting from non-optimal local-minima to optimal global-minimum quantum-tunneling!!! (2) “SHAZAM”: quantum-statistics “supersymmetry”- transmutation from Fermi-Dirac to Bose-Einstein 1/[+ 1/[e^(hw/kT) + 1] —-> 1/[e^(hw/kT) – 1] ~ 1/f power-spectrum, with no local-minima and permitting Bose-Einstein condensation( BEC) via a noise-induced phase-transition (NIT). Frohlich biological BEC & BNN 1/f “noise” power-spectrum concurred!!! Siegel’s work[]IBM Conf. Comp. & Maths.,Stanford(1986)] was used as “Page-Brin” Google first search-engine!!!

Commenting is closed for this article.