Reservoir Computing and Benchmarking of Neuromorphic Systems for Swap-Constrained Autonomous Processing
Abstract
Random projection networks are a class of learning algorithm, which use high-dimensional random projections of information as a basis for training simple linear classifiers. The advantage of this approach is the lite computational cost associated with learning and short training times. These advantages make random projection networks suitable candidates for implementing on-device learning systems for on-the-edge intelligence. In particular, we focus on a feed forward RPN known as the Extreme Learning Machine (ELM) as a baseline for spatial information, and two recurrent RPNs, known as the Liquid State Machine (LSM) and Echo State Network (ESN), for processing spatio-temporal information. The fist focus of this work was to investigate advancements at an algorithmic level to make the algorithms more computationally powerful and hardware friendly. Secondly, we developed several baseline architectures for stochastic, digital, and analog implementations for size, weight, and power (SWaP) constrained platforms with on-device learning.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 26, 2019
- Accession Number
- AD1079216
Entities
People
- Abdullah Zyarah
- Dan Christiani
- Dhireesha Kudithipudi
- Humza Syed
- Nicholas Soures
- Qutaiba Saleh
- Ryan Bruell
- Swatika Ramakrishnan
Organizations
- Rochester Institute of Technology