Brain Inspired Next Generation Deep Learning:Efficient and Persistent Online Learning with Spikes
Abstract
This project aims at fundamental advances over the state-of-the-art in deeplearning (DL) for neuromorphic spiking neural networks (SNNs) by: i) offeringguaranteed uniform convergence to user-specified learning objectives in a singlepass through the data; and ii) operating entirely on locally available informationdistributed across the SNN architecture, directly amenable to efficient implementation on reconfigurable large-scale neuromorphic computing platforms. Exact incremental and decremental on-line algorithms for supervised, unsupervised, and reinforcement learning for SNNs will be investigated based on information theoretic models of spike-timing and postsynaptic potential dependent eligibility for local synaptic plasticity towards global data-driven objectives. Efficient implementation of these algorithms for persistent online learning in object identification and classification will be demonstrated on massively parallel neuromorphic hardware using streaming spike train data from event-driven vision and audition sensors, as well as standard and internal benchmark datasets of interest to Navy operations through a partnership with NIWC Pacific. We anticipate widespread use and further development by the neuromorphic community will support a broad range of applications of persistent lifelong learning for pervasive autonomy in the internet of things (IoT).
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 08, 2020
- Source ID
- N000142012405
Entities
People
- Gert Cauwenberghs
Organizations
- Office of Naval Research
- United States Navy
- University of California, San Diego