Learning on the Edge With Hyperdimensional Computing
PI: Vasant Honavar (Informatics and Intelligent Systems)
The PI will be responsible for covering tuition for the graduate student.
Deep learning has become the methods of choice for a broad range of applications that involve learning predictive models from large and complex data sets. However, deep learning techniques are computationally expensive, include many tunable hyperparameters, are data hungry. Their computational needs make them impractical for learning from data in resource constrained environments, e.g., edge computing, wearables, IoT sensors, or mobile platforms. with limited computational capabilities and power, bandwidth or connectivity constraints. Hyperdimensional (HD) computing offers a promising computationally efficient alternative to deep learning methods in such settings. This project aims to develop and evaluate lightweight, HD computing based machine learning framework for learning on the edge, that ls, learning predictive models from data being acquired by edge devices. The resulting methods will also help significantly reduce the carbon footprint of machine learning.
The research is guided by three key questions:
RQ1: How does HD Computing compare to deep learning and classical baselines in terms of learning efficiency and accuracy on edge devices?
RQ2: What algorithmic modifications are necessary to adapt HD computing models to accommodate device-specific constraints (e.g., memory, precision, latency)?
RQ3: Can HD computing-based machine learning algorithms be generalized to cope with different streaming data scenarios and hardware platforms?
Project Objectives:
● Design and evaluate multiple HD-based learning algorithms tailored for edge environments
● Benchmark performance and energy efficiency against traditional models
● Deploy models on selected edge devices and validate using real-time or simulated streaming data
Long-Term Goal:
To establish a modular and deployable HD computing framework that enables efficient learning on edge devices, with the potential for expansion into embedded AI systems, real-time analytics, and low-power settings for applications such as environmental monitoring, health monitoring (using wearable sensors).
Connection to ICDS Mission:
This work directly supports the ICDS mission by advancing novel, computation-efficient methods in machine learning and applying them to real-world settings at the intersection of data science, embedded systems, and artificial intelligence. By fostering interdisciplinary collaboration across computer engineering, cognitive computing, and edge AI, the project aims to produce scalable and impactful innovations in machine learning on the edge.