Xiaofan Yu introduces LifeHD, a system enabling unsupervised lifelong learning directly at the edge. Unlike traditional methods relying on offline data, LifeHD learns continuously from unlabeled streaming data in real time. This approach supports CoCoSys’s goal of developing cognitive-inspired solutions, demonstrating the potential of neural-symbolic algorithms in practical, dynamic environments, expanding the boundaries of hyperdimensional computing by applying it to challenging scenarios and promising applications like environmental monitoring and human activity recognition, enhancing adaptability and real-world impact.

1. Tell us about the findings of your recent work entitled “Lifelong Intelligence Beyond the Edge using Hyperdimensional Computing” and how it supports the goals of CoCoSys.

Traditional edge ML deployment involves a pipeline of offline data collection, offline ML model training and ML model compression on an embedded system. However, this pipeline might not work in reality as the offline data can differ significantly from real-world data. To address this issue, we design and deploy a system called LifeHD for on-device unsupervised lifelong learning at the edge. LifeHD adopts a different approach: we directly deploy an edge device into an unknown environment, without prior data collection, and let the system learn continuously from unlabeled streaming data in a “lifelong” manner.

LifeHD contributes to CoCoSys’s goal of developing cognitive-inspired system solutions for real-world applications, particularly through neural-symbolic algorithms. The achievements of LifeHD highlight the potential of neural-symbolic solutions in real-world applications that require unsupervised learning, real-time responses and efficient implementations.

2. How do your research findings push the boundaries of what we currently know or can do in the field of HD computing?

 LifeHD advances the existing boundaries of HDC by expanding it to the challenging unsupervised lifelong learning scenario and providing a real system implementation. While existing HDC applications have shown superb performances mainly in supervised tasks, LifeHD proves that HDC can effectively learn and store key patterns even in the challenging scenario of streaming, unlabeled data. The design of a short-term and long-term memory hierarchy is critical to achieve this goal.

3. What are some real-world applications or examples of your research that people might encounter in their daily lives? 

LifeHD is motivated by the challenges of deploying edge intelligence in real-world applications. For example, in environmental monitoring, one may want to deploy a sensor system into a natural reserve and track major environmental patterns. However, it is very common that collecting a large amount of high-quality data for a specific location is hard (e.g., the system is supposed to be deployed on a cliff). In this case, LifeHD offers a novel solution by allowing the system to be deployed directly into such locations and “learn” automatically. Our system begins by collecting data and automatically recognizing, differentiating, and memorizing major patterns. The learning outcomes are, for example, being able to successfully detect four different seasons in the environment, after continuous data collection and learning.
LifeHD is designed for general lifelong learning scenarios at the edge and can also be applied to other applications, such as human activities monitoring, sound recognition and image classification, as we demonstrated in the paper.

4. What inspired you to pursue this research, and why do you think it is important?

The objective of LifeHD is inspired by real-world applications, while its design methodology is inspired by the cognitive science of how humans learn. What I find particularly fascinating about this research is its connections to biological or human lifelong learning. Like humans, who are born knowing nothing and must learn about their environment without supervision and with limited memory resources, LifeHD aims to operate in a similar manner. The parallels between human lifelong learning and the LifeHD system are striking.

I personally really appreciate this echo with human intelligence, and I believe recognizing this connection is important for designing truly “intelligent” systems in the next generation, which work both individually and collaboratively with humans.