Air Force Research Laboratory/Progress Toward Neuro-memristive Processors
Prime Contractor, 2016-Present
Machine intelligence is a catalyzing technology with world-changing consequences. Intelligence is defined as “the ability to acquire and apply knowledge and skills.” In other words, intelligence is intimately related to (or is) learning. Learning is a memory- processing operation. Current digital computing methods separate memory and processing, resulting in very high 𝐶𝑉^2 losses. The discrepancy in SWaP efficiency between modern digital computing methods and biology is 4 to 9 orders of magnitude. By eliminating or reducing the energy associated with computing synaptic integration and adaptation through a reduction to an analog processes, neuromemristive processors offer the most efficient known path toward practical learning processors. Due to the inextricable link between intelligence, learning and memory-processing communication, the ‘end game’ for artificial intelligence occurs when a unified memory- processing architecture achieves state-of-the-art results on primary performance benchmarks. This is the goal of AHaH Computing.
We believe that AHaH Computing and kT-RAM and similar 'in-memory' structures offer the most practical and robust solution to achieving integrated machine intelligence. However, there are a number of technologies that will be enabled via a Back-End-Of-Line (BEOL) CMOS Memristor service. Examples include alternate neuromemristive processors, stacked memory-on-top processors, reconfigurable logic, multi-state logic and memory, adaptive filters, physically unclonable devices, and more. Indeed, research into memristive circuits is growing exponentially. Our objective in this program is therefore to create the tools to enable a memristor wafer services business while at the same time demonstrating the last remaining technical milestones of AHaH Computing before chip production.
Air Force Research Laboratory/Emerging Algorithms for Autonomous Sense-Making Operations
Prime Contractor, 2012-Present
Today’s computing information systems are continuously bombarded with sensor, machine, and user generated data. This data can originate and travel through multiple communication channels such as airways, wires, optical links, radar, radio, and/or cyber space. Thus, the main challenge information analysts face today is how to deal with the vast amounts of data available given that in order to process all this information would require beyond human abilities as clearly described in the recent Technology Horizons report by the Air Force chief scientist Dr. Dahm. “Although humans today remain more capable than machines for many tasks, natural human capacities are becoming increasingly mismatched to the enormous data volumes, processing capabilities, and decision speeds that technologies offer or demand; closer human-machine coupling and augmentation of human performance will become possible and essential”. As a result, this applied research topic seeks innovative ideas where emerging computing algorithms can be applied to intelligent information processing or autonomous sense making operations. The basic goal is to develop information software platforms capable of performing autonomous operations that would enhance the performance, operation, and decision making capabilities of the user by enabling autonomy to the system itself.
Air Force Research Laboratory/VLSI-Memristor Building Blocks for Future Autonomous Air Vehicles
Prime Contractor, 2011-Present
Future information processing systems will require some basic level of intelligence to aid user decision-making capabilities. Neuromorphic computing promises to allow the development of intelligent systems able to imitate natural neuro-biological processes. This is achieved by mimicking the highly parallelized computing architecture of the biological brain. In theory, neuromorphic computers are suitable for applications in decision making by being able to intelligently process vast amount of information in parallel and in real time. The continuing improvements within microelectronics research and development allow for the integration of hundreds of millions of CMOS transistors within a single silicon chip within just about 1 cm2. For example, the number of neurons in honey bee brain is equal to approximately 950,000, and honey bees can perform autonomous functions un-matched by even the powerful supercomputers and clever algorithms in the world today such as navigation and organization skills to find and store food. However, given today’s technology integration capabilities, it should be possible to recreate in hardware, the computing architecture that enables honeybees or any other simple biological organisms to perform autonomous functions. In addition, recent developments in memristor-based technologies claim the invention of the physical analog to the synapse. In theory, memristor-like technology (passive memory devices: memristor, magnetic junction, and/or continuous variable resistance devices) would enable massively parallel large scale neuromorphic computing processor architecture development.
Office of Naval Research/Chaos Computing
Prime Contractor, 2012-2013
Nonlinear dynamics has revealed a rich array of behaviors, especially those related to chaos including routes to chaos, high and low dimensional chaotic attractors, crises, transient chaos, and Hamiltonian strange kinetics. In neural systems, measured phenomena includes chaos, synchrony, and cascading avalanches demonstrating that information processing in the brain is not just anatomical, but also dynamical. This program seeks to take advantage of the richness of nonlinear dynamical systems and insights from neural systems to devise new approaches to computation. Possible approaches include utilizing computing with attractors, transient chaos in high dimensional systems, chaos controlled reconfigurable logic gates, and pattern formation. Cues may be taken from neuroscience with network topologies of excitatory and inhibitory connections and plasticity of learning through strengthening of connections at synapses. We seek a “plastic” computational network that can be programmed to adjust rapidly without a physical rewiring to seek optimal solutions to problems. The sought after approach is to include nonlinear dynamical behaviors into an information processing system that can optimize solutions to complex problems. This research promises a revolution in information processing for areas such as pattern recognition where a complex circuit can self-organize by morphing between logic gate configurations to search for specific patterns, such as, faces, or vehicles. The incorporation of concepts from neural cognitive behavior can led to feedback and self-organization designs to increase the effectiveness of information processing. Novel computing can allow for a versatile response to information flow which can lead to new paradigms for the optimization of solving complex problems, such as the control of robots and other autonomous systems.
DARPA/Physical Intelligence (PI)
SETA Consultant, 2009-2011
The Physical Intelligence program will address its overall objective through a coordinated effort in three complementary domains: theory, implementation, and analysis. The objective of the theory domain is to develop and validate a physical formalism that unifies and expands ideas from diverse domains such as evolution, thermodynamics, information, and computation. The objective of the implementation domain is to demonstrate the first human-engineered open thermodynamic systems that spontaneously evolve nontrivial “intelligent” behavior under thermodynamic pressure from their environment. The objective of the analysis domain is to develop analytical tools to support the development of human-engineered physically intelligent systems and to understand physical intelligence in the natural world.
DARPA/Systems of Neuromorphic Adaptive Plastic Scalable Electronics (SyNAPSE)
SETA Consultant 2008-2011
Current programmable machines are limited not only by their computational capacity, but also by an architecture requiring human-derived algorithms to describe and process information from their environment. In contrast, biological neural systems, such as a brain, autonomously process information in complex environments by automatically learning relevant and probabilistically stable features and associations. Since real-world problems generally have many variables and nearly infinite combinatorial complexity, neuromorphic electronic machines would be preferable in a host of applications. Useful and practical implementations, however, do not yet exist.
The vision for the Systems of Neuromorphic Adaptive Plastic Scalable Electronics (SyNAPSE) program is to develop electronic neuromorphic machine technology that scales to biological levels. SyNAPSE supports an unprecedented multidisciplinary approach coordinating aggressive technology development activities in the following areas: hardware, architecture, simulation, and environment.
The initial phase of SyNAPSE developed nanometer-scale electronic synaptic components capable of adapting connection strength between two neurons in a manner analogous to that seen in biological systems and simulated the utility of these synaptic components in core microcircuits that support the overall system architecture.
Continuing efforts will focus on hardware development through microcircuit development, fabrication process development, single chip system development, and multi-chip system development. In support of these hardware developments, SyNAPSE seeks to develop increasingly capable architecture and design tools, very large-scale computer simulations of the neuromorphic electronic systems to inform the designers and validate hardware prior to fabrication, and virtual environments for training and testing simulated and hardware neuromorphic systems.