PORTLAND, Ore. All the rage a decade ago, neural networks are making a comeback at the Defense Advanced Research Projects Agency (Darpa).
Last year, the agency resurrected efforts to create brain-like electronic circuitry to automatically learn the relevant features and associations needed to recognize objects. Last week, IBM Corp. received a $16.1 million contract from Darpa for a program called SyNAPSE (Systems of Neuromorphic Adaptive Plastic Scalable Electronics). The program aims to "break the programmable machine paradigm and define a new path forward for creating useful, intelligent machines," according to a Darpa announcement.
Current systems for pattern recognition of enemy tanks, for example, depend on hand-crafted algorithms, but SyNAPSE aims to automate pattern recognition by mimicking the human brain. According to program manager Todd Hylton, it also aims is to "to develop a electronic neuromorphic machine technology that scales to biological levels."
To achieve ithis goal, researchers need a synapse--the adaptive connection between neurons in the brain that changes values in response to perceived patterns. Synapses change the connection strength between neurons to learn new patterns without traditional programming.
IBM's charter is to develop materials that facilitates the changes that adapt to new patterns, categorize the features that will enable recognition and can interface with conventional computers to provide them with learning capabilities.
Besides developing a working synapse, the project will also simulate synaptic components in special-purpose cores that support the adptaive learning architecture needed for automatic pattern recognition systems.
If successful, follow-on funding will be provided to develop neuromorphic hardware systems, design tools and large-scale simulations intended to validate the feasibility of scaling synapse-based pattern recognition systems to levels approaching the human brain.
Other contracters for Darpa's neural network program include HRL Laboratories and Hewlett Packard Co.