N00014-19-S-SN08 - Office of Naval Research - Navy.mil Special...

N00014-19-S-SN08 - Office of Naval Research - Navy.mil


Special Notice N00014-19-S-SN08 Special Program Announcement for 2019 Office of Naval Research Basic and Applied Research Opportunity: “Science of Artificial Intelligence – Basic and Applied Research for the Naval Domain

Title: Brain-Inspired Deep Learning with Spiking Neurons (AI Fundamental Research)


Background: Deep Neural Networks have achieved strong performance in tasks such as image classification. However, they require significant time and computing resources to train. While the steady advance of Moore's Law has enabled their success, integrated circuit technology is reaching fundamental physical limits due to thermal and stability issues.
At the same time, there is a need in the Department of Defense (DoD) for advanced compute capabilities on platforms severely limited in Size, Weight, and Power. Another need is Machine Learning (ML) systems that can learn without access to large amounts of labeled training data.

The brain is a highly decentralized event-driven computing device, processing multiple asynchronous streams of sensorimotor data in real-time. Neurons communicate through spikes - brief impulses transmitted to other neurons through synapses. Experimental evidence shows that not only the rate of spike firing, but also the precise timing of spikes can be important for processing information

Deep-learning models solve problems by assuming static units that produce analog output, which describes the time-averaged firing-rate response of a neuron. These rate-based artificial neural networks (ANNs) are easily differentiated, and therefore can be efficiently trained using stochastic gradient descent learning. The recent success of deep learning demonstrates the computational potential of trainable, hierarchical distributed architectures.

Simulations of deep learning networks is highly computation intensive, which consumes power and limits the efficiency of mobile devices. Neuromorphic hardware based on spike communication between chips is a thousand times more energy efficient and more compact than digital chips. Because spikes are discontinuous, which precludes computing gradients; it has not been possible to use stochastic gradient descent to solve complex problems

. Research in spike- based computation has been impeded by the lack of efficient supervised learning algorithm for spiking networks.

Objective:

The goal of this project is to develop new learning algorithms for spiking neurons that will allow deep learning spiking networks to be built that can solve complex real-world problems. The consequences would be far reaching in terms of both the practical applications and the theoretical insights into how to compute with spikes.

Nature is an existence proof that this is a solvable problem.