A Roadmap to Neuromorphic Systems Proposed

 Neuromorphic Computers  

Microprocessors configured more like brains than traditional chips could soon make computers much more powerful and energy efficient. Now researchers have created a roadmap for the development of brain-like neuromorphic systems that may serve as a guide for development.

Cognitive computers, artificial neural networks, neuromorphic systems, and similar efforts to cast the intelligence of the human brain into silicon chips is an area of rampant development in major corporations, from IBM to Qualcomm, and at major government-sponsored efforts, including DARPA's Systems of Neuromorphic Adaptive Plastic Scalable Electronics Human Brain Project (SyNAPSE program).

Now researchers at Georgia Tech have created a roadmap for the development of brain-like neuromorphic systems that may serve as a guide for development.

According to the authors' abstract:

Neuromorphic systems are gaining increasing importance in an era where CMOS digital computing techniques are reaching physical limits. These silicon systems mimic extremely energy efficient neural computing structures, potentially both for solving engineering applications as well as understanding neural computation. Toward this end, the authors provide a glimpse at what the technology evolution roadmap looks like for these systems so that Neuromorphic engineers may gain the same benefit of anticipation and foresight that IC designers gained from Moore’s law many years ago. Scaling of energy efficiency, performance, and size will be discussed as well as how the implementation and application space of Neuromorphic systems are expected to evolve over time.

"If you look at the technology we have now, and forward to that which will soon be available -- for instance, 11 nanometer processes -- then you come to the conclusion that eventually we really could build a human-like cortex in something that could use as little as 50 watts of power and be small enough to put on your desk."

"If you look at the technology we have now, and forward to that which will soon be available -- for instance, 11 nanometer processes -- then you come to the conclusion that eventually we really could build a human-like cortex in something that could use as little as 50 watts of power and be small enough to put on your desk," Georgia Tech professor Jennifer Hasler  told EETime's R. Colin Johnson.

The roadmap is authored by Hasler and Georgia Tech doctoral candidate Harry Bo Marr, now graduated and working at Raytheon as the technical lead of DARPA's Arrays at Commercial Timescales (ACT) program and lead electronic warfare digital architect.

Hasler and Marr's roadmap also relied upon results obtained by Georgia Tech doctoral candidate Suma George, who experimentally demonstrated, for the first time, that neuromorphic systems based on circuits that tightly model biological principles have clear advantages over other approaches, including typical analog signal processing techniques.

Transistor based models of neurobiological computation - Hasler and Marr

Hasler's own research emphasizes the key role that properly designed analog processing elements will have on neuromorphic systems, specifically field-programmable analog arrays (FPAAs) that she and colleagues at Georgia Tech have been perfecting for several years.

FPAAs are already available from Anadigm.

FPAAs are similar to field-programmable gate arrays (FPGAs) but include reconfigurable analog elements. FPAAs are commercially available from Mesa, Az.-based Anadigm, but Hasler claims Georgia Tech's FGAAs "dwarf the programmability and capability of the Anadigm components" by virtue of housing "hundreds of thousands of programmable parameters, enabling them to be used for system-level computing, not just analog glue logic."

Field-programmable analog array (FPAA) Board - Image Source: Rob Felt / Georgia Tech

Related articles Dennis Hong On Robot Evolution and The DARPA Robotics Challenge  Researchers Develop Neuromorphic Chip With Cognitive Abilities  The Human Brain Project Has Begun Building the Super Brain

Hasler believes the clear path to desktop neuromorphic systems will require such analog system-on-chip (SoC) approaches to computation in order to achieve the low-power devices necessary to emulate billions of brain-like neurons connected by trillions of learning synapses. In all, she predicts that desktop neuromorphic systems that rival the compactness of the human brain will require "eight orders-of-magnitude" (100 million times) reduction in power over the digital supercomputers simulating them today.

In Hasler and Marr's roadmap they describe the specific computational milestones that have already been achieved -- such as the single-transistor synapse and the FPAA -- as well as the algorithms and computational models that have been proven out so far. The roadmap then proceeds to detail the areal density of artificial neurons and synapses that will be necessary to realize a low-power neuromorphic system whose size rivals that of a human brain.

"Useful neural computation machines based on biological principles at the size of the human brain seems technically within our grasp."

Finally, the roadmap addresses the software tools that will be necessary to design these neuromorphic chips as well as the learning techniques, network topologies, novel interconnection devices -- such as memristors -- and how these components could be developed into usable chip arrays. For instance, memristors could serve for slow-scale modulatory parameters in neuromorphic systems.

Throughout, the roadmap emphasizes a modular approach, such as successfully emulating one layer of a human brain cortex before attempting multilayer devices, as well as the major engineering hurdles that need to be surmounted, such as using local-interconnection techniques to reduce the complexity of communications traffic among billions of neurons and trillions of synapses. The paper's overall conclusion is that "useful neural computation machines based on biological principles at the size of the human brain seems technically within our grasp."

It is worth noting that Gill Pratt is heading up the DARPA SyNAPSE project.  Pratt also heads the DARPA Robotics Challenge, so it is very evident that neuromorphic chips and robotics are two converging projects with substantial impacts for the future.

SOURCE  EE Times

By 33rd Square

Embed