Image

Guangyu Xu, associate professor of electrical and computer engineering and adjunct associate professor of biomedical engineering at UMass Amherst, has been awarded a three-year, $425,000 grant from the U.S. National Science Foundation (NSF) to develop an optoelectronic artificial intelligence (AI) system that can efficiently learn from and send feedback to biological circuits, essentially creating a high-performing hybrid intelligence framework. This fundamental research may ultimately push forward the boundaries of AI hardware, human-machine co-learning, brain-computer interfacing and disease modeling.

Xu will collaborate with Professor Sebastiaan van Dijken at Aalto University to establish optogenetic-neuromorphic (using light to control artificial neural networks inspired by the brain) co-designed cell interfaces, leveraging their expertise on optical control over cell activities and bio-inspired photodetection. Their joint effort is likely to offer closed-loop, energy-efficient, and high-accuracy cell interfacing, the functionality of which will be showcased in biological neural networks.

“Knowing how neurons communicate, namely neural decoding, is a very data-intensive and latency-sensitive task, because you want to know what they are doing collectively within milliseconds,” says Xu. “Optoelectronic neuromorphic devices lend themselves to a fundamentally new method of neural decoding.”

Leveraging the memristive readout from engineered semiconductors, this novel method will exclude superfluous data while maintaining biological insight. As a result, the researchers can study neural activities in a highly efficient manner.

Taking it one step further, the researchers aim to build artificial neural networks (NN) based on biological data collected from their neuromorphic hardware to determine whether cells are in their normal or abnormal states. “This is a key step in a hybrid intelligence framework, where the biological computing input is directly applied to train AI-based machine learning models,” says Xu. 

The output of these AI models have the capability to inform the neuromodulation hardware (e.g. close-packed LED arrays) on how to efficiently modulate biological circuits to the targeted state. “Such physical feedback closes the loop of our hybrid intelligence system, and provides us insight on finding neurological interventions, such as cooling down overly active cells,” he says. 

This research aims to leverage the benefits of both approaches: the computational power and programmability of traditional computing systems, alongside the adaptability, energy efficiency, and inherent learning capabilities of biological systems like the human brain. Their team efforts could help define the future trajectory of bioinspired and biomimetic AI technologies.