CENTER OF EXCELLENCE (COE): Extreme Neuromorphic Materials and Computing
Funding Agency:
- Department of Defense
The human brain’s cognitive functions emerge from the collective processing capability of simple computing elements, i.e., neurons, synapses, and dendrites. In such biological neural networks, spike sequences carry both spatial and temporal information for communication and processing. Moreover, neurons operate asynchronously in an event-driven manner, and biological neural networks demonstrate very energy-efficient information processing. Such biological neural networks have initially been simulated via software-only approaches, and the scale of simulated networks usually is small due to high communication overhead and limited parallelizable computing with conventional hardware. The artificial neural network (ANN) approach loosely models neuron functionality and the massive connection of neurons in a biological brain, but ANN ignores a lot of essential features of biological neural networks. Such a simplification makes the training process quite subtle, inefficient, and sensitive. Although ANNs have obtained substantial successes in many applications such as image and speech recognition, the incredibly high computing cost required by training and poor support of in-hardware learning emerge as the main obstacles. Neuromorphic models differ from their ANN counterparts in that they encode information via the temporal activation of neurons and precise emission of spikes. Over the past few years, many research efforts have been devoted to neuromorphic models to harness their higher computational potential. These efforts have focused on developing algorithms such as Widrow-Hoff type supervised plasticity rules for precise temporal spike-train recognition, back-propagation approximations for powerful data-driven learning, and spike-timingdependent plasticity (STDP) for scalable temporal learning in progressively deeper neural architectures. CMOS based hardware has been built to implement the neuromorphic models and shown limited architectural design perspectives. The most visible examples are IBM’s TrueNorth neurosynaptic processor and Intel’s Loihi neuromorphic chip. CMOS devices and circuits, as the building elements for these hardware systems, were not created or optimized for neuromorphic computing purposes in the first place.
Award Ceiling: $5,000,000
Pre-proposal inquiries and questions must be received in writing by electronic mail not later than 29
June 2023 at 11:59 PM Eastern Time (ET) to be considered.
White papers are highly encouraged and must be submitted electronically at
https://community.apan.org/wg/afosr/p/submitawhitepaper by 20 July 2023 at 11:59 PM
Proposals must be received electronically through Grants.gov by 24 August 2023 at 11:59 PM Eastern Time to be considered.
Valencia Thornton Contract Specialist Phone 703-696-7337
valencia.thornton@us.af.mil