Hyperdimensional Neural Computation for Real-Time Cognitive Learning

With each passing year, the volume of analyzable data continues to explode, with new information gathered from a variety of video, sensors, and informatics platforms. Therefore, we face increasing needs for machine learning techniques to adaptively analyze this large-scale data and transform that into actionable knowledge. Today's machine learning algorithms, however, have the following key technical challenges: (1) they are extremely slow and inefficient on lightweight embedded devices, e.g., smartphones or smartwatches, (2) they are very vulnerable to noise and failures that often exist in highly scaled technology and network, and (3) they lack brain-like cognitive support and adaptation to provide a high quality of learning and a reason for each prediction and decision.

To achieve real-time performance with high energy efficiency and robustness, we redesign algorithms using strategies that more closely model the human brain. We leverage Hyper-Dimensional Computing (HDC), a brain-inspired method, motivated by the observation that the human brain operates on high-dimensional space. In HDC, objects are encoded into high-dimensional vectors to represent neural patterns with thousands of elements.

This encoding transforms data into knowledge that enables lightweight cognitive learning. In this project, we develop an infrastructure that supports a wide range of learning and cognitive tasks with high robustness and efficiency. Our open-source infrastructure provides real-time and dynamic learning, with a wide range of applications in intelligent healthcare, environmental monitoring, and smart distributed systems.

The project will also support students through synergistic outreach plans and educational activities, including programs for K-12 students, undergraduate research opportunities, and new course development.

The novel research approaches introduced in this project aim to lay the foundations for deeper integration of brain-inspired mathematics, learning algorithm, and hardware. This includes the development of: (1) encoding methods that map various spatial-temporal data into high-dimensional space, including simple numerical values to complex video/images, (2) algorithmic solutions that enable brain-like learning and cognitive tasks over encoded data, and (3) hardware-software libraries that significantly accelerate the HDC algorithms.

Our programmable processor natively supports HDC operations while utilizing extensive parallelism offered by multiple hardware platforms. We will evaluate the effectiveness of our framework on multiple large-scale systems, including smart home and environmental monitoring. Our prototypes will be fully released under an established open-source library for wide public dissemination.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.