AI-driven Intelligent Computing Enabled Brain-computer Interface

Active Sites

Duke University


Brain-computer interface (BCI) is used as a case study to determine how AI-drive intelligent computing can substantially contribute to multiple, critical tasks for neural signal processing, and facilitate the building of a robust, real-time BCI platform for human usage. The objective of a BCI system is to create alternative communication pathways for human patients to interact with the environment by prosthetics and/or wheelchairs when their normal motor function is impaired by amputation, trauma or disease. In these cases, the controllability of a patient is substantially limited by the impairments; e.g., a patient with an injury to the spinal cord is not able to control a wheelchair by hand. Hence, it is extremely important to develop a BCI system where neural signals from the brain are sensed, amplified, filtered, transmitted/received and decoded to control a prosthetic limb, an electric wheelchair, or other external devices. A typical BCI platform consists of three major components: (1) a sensing system implemented with analog and radio-frequency (RF) circuits that measure, amplify, and digitize neural signals; (2) a signal processing unit implemented with digital circuits that process and decode neural signals to accomplish a particular task; and (3) a programmable unit for system control and reconfiguration. An AI-driven intelligent computing may well serve the above three components from multiple aspects.


Our main goal is to develop an AI-driven intelligent computing system that can (i) achieve a reliable low-power implementation of many analog/RF circuit blocks, (ii) optimize the adaptive-learning process that is needed by two major steps of the neural signal processing flow – feature extraction and movement decoding, and (iii) enable a programmable unit with “self-evolving” feature in the BCI system for control and reconfiguration.


Xin Li and Mary Cummings (ECE, Duke) will be responsible for the development of computing algorithm and coordination with hardware systems. Miroslav Pajic and Hai Li (ECE, Duke) will be responsible for the hardware-level development and system integration.

Experimental Plan and Industrial Relevance

(i) We will design an AI-driven intelligent computing platform to facilitate a self-evolving platform that takes on-chip sensor data as input and self-heal the analog/RF circuits to achieve improved robustness and reduced power consumption for the required analog/RF operations.

(ii) Decoding the intension of the user from his/her brain signals requires the implementation of a self-evolving platform for adaptive learning.

During feature extraction, the stochastic time-domain brain signals are converted to the frequency domain by calculating the power spectral densities (i.e., the BCI features). Power modulation in specific frequency bands have been well demonstrated to strongly correlate with brain kinetic parameters by many neuroscience experiments and power spectral density estimation requires learning a statistical model (e.g., autoregressive model) to characterize the brain signals. The learning process must be repeated every few seconds so that the time-varying behavior of brain can be accurately tracked. Once the feature extraction is completed, an analytical mapping between the BCI features and the output movement parameters (e.g., direction, velocity and acceleration) must be adaptively learned to control the external devices such as wheelchairs and/or prosthetic limbs. An adaptive learning process will be developed to solve this problem (which can be formulated as either a classification or regression task) for reliable long-term usage

(iii) In our application, system reconfiguration is needed in order to change application scenarios (e.g., computer cursor control versus wheelchair navigation) and/or update decoding algorithms from time to time.

It is important to emphasize that once the BCI system is re-programmed by software, the proposed AI-driven intelligent computing will adaptively fit the new program and optimally accomplish the required computing tasks with high reliability and low power consumption. The “self-evolving” feature of AI-driving intelligent computing, hence, will be designed.

While the obvious application of our work is in healthcare, the proposed research will be of interests to multiple groups of industrial partners, including defense and transportation industries.


The first-year deliverables include an AI-driven intelligence computing system for the reliable low-power implementation of analog/RF circuit blocks. The end-of-project deliverables include (i) an AI-driven intelligence computing system to enhance the efficiency and robustness of analog/RF circuits, decoding process with “self-evolving” feature; and (ii) the software deployment and hardware prototype.

Milestones and Time-to-Completion

The estimated duration of this project is 3 years. The milestones are listed in the following table.

Year 1

Year 2

Year 3

Design AI-driven intelligence computing system for analog/RF circuit

Develop the intelligent system for the decoding process

Develop “self-evolving” feature of the intelligent system

Number of Graduate Students Supported




Total Cost to Completion