Efficient Probabilistic Computing with Analog Content Addressable Memory Systems
Abstract
Deep neural networks are of limited value for a range of applications where inspectability and explainability are critical, training data may be limited, and/or where domain knowledge and historical expertise needs to be incorporated in critical decisions. Examples include surveillance and robotics, cybersecurity systems, decision making and situational intelligence with multimodal inputs (medical, autonomous driving systems), and global system modelling (ground water models, smart factories, robotic/AUV fleet deployments) for risk/failure assessments and predictive maintenance. In these domains, important probabilistic techniques include Markov chain Monte Carlo (MCMC) and Bayesian Inference which inherently require sampling and computing with probability distributions while managing models and priors. In addition, tree-based ensemble methods (Random Forests, extreme gradient boost trees, etc.) are popular machine learning approaches in similar domains as they are simple to train, do well with small data sets, and maintain reasonable interpretability for domain experts to verify and understand. This proposal aims to explore a novel approach to hardware acceleration of algorithms involving MCMC/Bayesian Inference as well as Random Forests/Decision Trees. The project leverages 1) recent work showing efficient acceleration of finite automata (state machines), augmented with 2) a newly invented analog content addressable memory (a-CAM) that can search and operate on continuous values and intervals, substantially improving data density to reduce power and area. We propose a non-von Neumann computing system where costly data movement is minimized and more efficient data storage is enabled with the use of non-volatile technologies.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Oct 01, 2019
- Source ID
- W911NF1910494
Entities
People
- John Strachan
Organizations
- Army Contracting Command
- Hewlett Packard Enterprise
- National Security Agency