Brain-Inspired Hyperdimensional System for Robust, Efficient, and Real-Time Learning

Abstract

To achieve real-time performance with high energy efficiency and robustness, we exploit HyperDimensional Computing (HDC) as an alternative computing method that mimics important brain functionalities towards high-efficiency and noise-tolerant computation. The fundamental novelty of this project is to design an end-to-end framework that systematically combines HDC with existing well-developed algorithms to significantly enhance efficiency, robustness, and learnability. Unlike prior research that looked at HDC mainly as a classifier, our framework, called DeepHD, expands HDC functionality to learning, cognitive, and computing. Our solution targets supervised/unsupervised learning and optimization applications. (1) We propose novel algorithmic innovations focusing on regression, reinforcement learning, clustering, and classification. DeepHD combines HDC learning capability to enhance the efficiency and robustnessof a wide range of algorithms. (2) We use DeepHD to inherently support cognitive functionalities related to memorization and association. The cognitive functionality enables today s algorithms to not only provide a higher quality of learning but also to reason about each prediction or decision. (3) We expand DeepHD application to the computing area by supporting stochastic arithmetic over HDCvectors. We exploit this functionality to accelerate the back-bone algorithm using uniform HDC data types and primitives. We accordingly design a programmable cognitive processor that natively supports HDC operations, utilizing extensive parallelism offered by FPGA, ASIC, and processing in-memory (PIM) platforms. (4) we design a scalable learning framework to distribute DeepHD computation over edge devices in IoT network, beyond transitional edge servers. We propose hierarchy-aware learning that makes IoT nodes capable of real-time learning from labeled/unlabeled data. We also utilize the idea of active learning to select the most informative samplesto communicate, encode, and create HD models based on minimal data. (5) We validate DeepHD by practical deployment on multiple large-scale IoT systems, which will verify both the quality of the proposed algorithms and the efficiency of system designs, ranging from the in-lab evaluation to deployment across a large scale mobile and stationary network. The project abstract is public.

Document Details

Document Type
DoD Grant Award
Publication Date
Apr 06, 2021
Source ID
N000142112225

Entities

People

  • Mohsen Imani

Organizations

  • Office of Naval Research
  • United States Navy
  • University of California, Irvine

Tags

Fields of Study

  • Computer science

Readers

  • Critical Infrastructure Protection in CBRN and WMD Threats.
  • Neural Network Machine Learning.
  • Parallel and Distributed Computing.

Technology Areas

  • 5G
  • 5G - Internet of Things
  • AI & ML
  • AI & ML - Neural Networks