Skip to main content

MohsenImanireport from IoT Analytics predicts that by 2025, there will be more than 30 billion Internet of Things (IoT) connections worldwide, up from 11.7 billion in 2020. Similarly, the market for wearable devices continues to grow, as noted in a forecast from Gartner: “The rise in remote work and increased interest in health monitoring during the COVID-19 pandemic was a significant factor driving market growth.” With this explosion in the number of IoT connections, and the growing demand for resource-constrained devices that can support everything from work-related activities to health monitoring, comes the need for increased efficiency, reliability and security. This is where the research of Assistant Professor of Computer Science Mohsen Imani comes into play.

Imani, who joined UCI’s Donald Bren School of Information and Computer Sciences (ICS) in July 2020, is director of the Bio-Inspired Architecture and Systems Laboratory (BIASLab). His group is working on a wide range of problems in the areas of brain-inspired computing, machine learning and embedded systems. “Our research goal is to design real-time, robust and transparent cognitive learning platforms that closely mimic brain properties,” says Imani. “We are also designing a secure and scalable learning framework for distributed learning/computing over a swarm of devices in IoT systems.”

Based on this work, Imani recently received two grants totaling more than $400,000. He will receive $310,000 from Semiconductor Research Corp. (SRC) for a multiyear project, “A Hyperdimensional Learning System for Efficient, Robust and Secure Online Learning.” He will also receive $92,000 from Cisco for the development of EdgeHD, a brain-inspired hyperdimensional computing system.

Research Into Brain-Inspired Computing
As a Ph.D. student at UC San Diego, Imani started designing hyperdimensional computing algorithms that can work like a brain. “It was a new kind of algorithm that tries to mimic the human brain at an abstract functionality level,” says Imani. Hyperdimensional computing uses mathematics to approximate and explain how the human memory works, and Imani and his group are trying to move beyond research into the cognitive learning domain, expanding into classification, memorization and reasoning tasks. The work requires an interdisciplinary approach, which is one of the main reasons Imani says he came to UCI.

“It was because of the faculty and breadth of interdisciplinary research here, allowing me to expand my work into different directions,” he says. “[My research] is close to neuroscience, because we are continuously looking at more accurate models for the human brain. It’s close to cognitive science, because we are looking at how to design the next generation of robots and devices with cognitive capabilities,” he continues. “It’s connected to computer science, as we are developing new algorithms and processors that can mimic brain properties, so it’s actual hardware design.”

A Hyperdimensional Learning System
The SRC project aims to enable brain-like cognitive learning to address real-world problems, such as classification tasks. In particular, working closely with several companies, including Intel, IBM and NXP, Imani and his team will design a hyperdimensional processing unit (HPU) that

  • uses regenerative and dynamic encoding strategies to map various data types to neural activity;
  • enables classification tasks over encoded data; and
  • develops a private distributed learning framework.

“The main idea is about using hyperdimensional computing to enable learning on distributed systems,” says Imani. “As a result, our personal battery-based devices, such as our smartwatches and smartphones, will learn more efficiently than ever before and will securely adapt to new information without needing lengthy training times and large-scale data centers.” Because the data is encoded into high-dimensional space, that encoding helps keep the data private and secure.

A Neural-Inspired EdgeHD System
The CISCO project, “EdgeHD: Brain-Inspired Hyperdimensional Computing for Efficient Robust, and Scalable Edge Learning,” represents a breakthrough in edge computing as Imani and his team work to develop a neural-inspired hyperdimensional system to support various learning and cognitive tasks. For example, EdgeHD aims to support prediction tasks based on local learning. “We’re designing a system that can implement different algorithms on various small devices with very little power, battery-based devices,” says Imani.

The project is also focused on increasing robustness. “If you’re running these algorithms on hardware, they may have a lot of errors,” says Imani, “but our algorithm is inherently robust to noise and failure.” EdgeHD should lead to more efficient and highly scalable learning, allowing Zettabytes of data to be analyzed at the source and allowing IoT systems to adapt to new information without requiring time-consuming training.

Both projects aim to reduce reliance on large data centers with high power consumption. “Our goal,” says Imani, “is to enable efficient, robust and secure learning on small embedded devices.”

— Shani Murray