Reconfigurable Chip Aims to Give AI the Same "Lifelong Learning" Capabilities as the Human Brain

A team of scientists from a number of universities and laboratories across the US have come up with what they believe is a means to encode the same kind of lifelong learning ability as enjoyed by the human brain into computer hardware — considerably boosting the potential for future artificial intelligence systems.

"The brains of living beings can continuously learn throughout their lifespan," explains Shriram Ramanathan, professor in Purdue University's School of Materials Engineering, of the breakthrough. "We have now created an artificial platform for machines to learn throughout their lifespan."

The idea: Ditching fixed-function computer chips, as commonly found in machine learning systems today, and replacing them with reconfigurable devices, which can discard previous links and forge new ones — mimicking the human brain's ability to learn over time.

"If we want to build a computer or a machine that is inspired by the brain," Ramanathan explains, "then correspondingly, we want to have the ability to continuously program, reprogram and change the chip."

Built on perovskite nickelate electronics, the proof-of-concept hardware developed by the team can be programmed on-demand using electrical pulses — causing hydrogen ions to relocate and changing the functionality of the device between acting like resistor, memory capacitor, a neuron, or a synapse depending on where the hydrogen is concentrated.

To showcase the concept's potential, the team worked on a use-case in which a neural network takes advantage of the reconfigurability of the device to improve its capabilities at recognizing patterns in electrocardiogram data or images of digits.

"We demonstrated that this device is very robust," adds Michael Park, co-author of the paper. "After programming the device over a million cycles, the reconfiguration of all functions is remarkably reproducible." As an added bonus, the hardware is produced using techniques entirely compatible with today's semiconductor manufacturing fabs — and operates at room temperature, without exotic cooling systems.

The team's work has been published in the journal Science under closed-access terms; thus far, a route to commercialization has not been disclosed.