MULTI-LAYER SELF-PROGRAMMING NEUROMORPHIC INTEGRATED CIRCUITS FOR DEEP LEARNING
Abstract
Technical Challenge: Computers have led to artificial intelligent systems that simulate the deep learning function in the human brain. However deep-learning in supercomputers requires significant human programing efforts and consumes tremendous energy, which fundamentally restrains supercomputers from deep-learning efficiently in real-time, and limits the developments of intelligent systems such as self-piloted unmanned aerial vehicles (SPUAVs) with self-control in erratic complex environments. By contrast, the brain performs deep-learning in analog parallel mode via multilevel neural networks, and facilitates real-time learning with an energy efficiency more than five orders of magnitude superior to that of the supercomputers. Research Approach: We propose to develop a multi-layer self-programming neuromorphic integrated circuit (MLSNIC) based on synaptic resistors and neuristors developed in our lab to emulate the deep-learning functions of the neural networks. With their parallel signal processing and real-time learning functions, the neuromorphic networks can self-program their algorithms, and operate with an energy efficiency more than six orders of magnitude superior to that of the supercomputers. To achieve the goal, we will (1) integrate the devices in a large-scale MLSNIC to emulate a biological neuronal network with its critical functions such as high-speed parallel signal processing, hierarchical self-programming, and real-time deep learning with a low-power consumption; (2) develop a general theoretical platform to establish MLSNIC; (3) benchmark the intelligent functions of MLSNIC in autonomous control, real-time deep-learning, and selfprogramming to navigate a SPUAV in erratic complex environments. Impact on DoD Capabilities: The intelligent neuromorphic network will find broad applications in military unmanned and manned systems requiring autonomous operation, control, awareness, decision making, and optimization in erratic complex environments. “Multi-Layer Self-Programming Neuromorphic Integrated Circuits for Deep Learning”
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 12, 2021
- Source ID
- FA95502010230
Entities
People
- Yong Chen
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of California, Los Angeles