Nieke Roos
23 March

Researchers from Eindhoven University of Technology want to build a device where brain cells work together with silicon microchips. The Bayesbrain project, funded by EAISI, aims to create the world’s first brain-on-chip computer to solve real-world problems in real-time, using limited amounts of energy. The scientists, led by Regina Luttge and Bert de Vries, believe that, in the future, such a hybrid device could enhance ultra-low-power wearables, IoT devices or controllers with AI technology.

Besides being the most intelligent computing system ever, the brain is also very energy efficient. It uses orders of magnitude less energy than traditional computers, making it attractive for future sustainable computing hardware. At TUE’s Mechanical Engineering department, Luttge works on growing brain cells in electronic devices. De Vries is a researcher at the Electrical Engineering department, focusing on computational neuroscience, signal processing and machine learning. In Bayesbrain, they’ve joined forces to combine the power of the brain with that of the microprocessor.

csm_BayesBrain BvOF 2022_0224_AIJ BayesBrain Regina Luttge Bert de Vries_da0e276713
TUE researchers Regina Luttge and Bert de Vries have joined forces to combine the power of the brain with that of the microprocessor. Credit: Bart van Overbeeke

“The inescapable truth is that our modern computers consume too much power, but brain cells use orders of magnitude less power,” comments Luttge. De Vries adds, “We can’t achieve the same orders of magnitude change with current computers. To attain this, we need a paradigm switch, and hybrid computing involving brain cells could be the answer.”

The researchers are going to place about 1,000 brain cells in a microfluidic brain-on-chip device. Thanks to the compartmentalized microfluidic system, the cells will form tiny neural circuits. To keep them alive, they’ll be supplied with water, nutrients and an incubating environment. “They should form sufficiently mature neural networks after about three weeks. Once these networks form and reform, the cells will be ready to ‘talk’ to the silicon-based computer,” explains Luttge.

The team will start with a 100 percent silicon-based Bayesian computer to solve the inverted pendulum problem. This is a simple real-time control task, equivalent to balancing a stick on a moving platform. It’s a classic problem used in reinforcement learning, a machine learning training method that rewards favored outcomes and penalizes unfavored ones. “Once the silicon side of the device is balancing the inverted pendulum, we’re going to transfer some of the computational load to the brain cell side,” says De Vries. “We’ll start by replacing a small part of the load at first and slowly replace more and more of the silicon part with brain cells.”

This approach is dependent on the development of an interface that allows the brain cells to talk to the silicon-based computer and vice versa. “The communication interface is perhaps the greatest problem to be solved in this project,” Luttge points out. “Without it, the brain cells won’t be able to share the computational load.”