The brain’s secret to lifelong learning can now come as hardware for artificial intelligence

WEST LAFAYETTE, India – When the human brain learns something new, it adapts. But when artificial intelligence learns something new, it tends to forget information it has already learned.

As companies use more and more data to improve how AI recognizes images, learns languages, and performs other complex tasks, an article published in Science this week shows how computer chips can be dynamically reconfigured to accept new data. as the brain does, helps AI continue to learn over time.

“The brains of living beings can learn continuously throughout life. We have now created an artificial platform for machines to learn throughout their lives, ”said Shriram Ramanathan, a professor at Purdue University’s School of Materials, which specializes in discovering how materials can mimic the brain to improve computing.


Shriram Ramanathan, a professor of materials engineering with Purdue, is exploring ways to introduce artificial intelligence directly into equipment. (Photo by Purdue University / Rebecca McElho) Download image

Unlike the brain, which is constantly forming new connections between neurons for learning, the circuits in the computer chip do not change. The circuit that the machine has been using for many years is no different from the circuit that was originally built for the machine in the factory.

This is a problem in order to make AI more portable, for example, for autonomous vehicles or robots in space that will have to make their own decisions in isolated environments. If AI could be built directly into hardware rather than just running on software, as AI normally does, these machines could work more efficiently.

In this study, Ramanathan and his team built a new piece of equipment that can be reprogrammed on demand using electrical pulses. Ramanathan believes that such adaptability will allow the device to perform all the functions necessary to create a computer inspired by the brain.

“If we want to create a computer or a machine inspired by the brain, then, accordingly, we want to be able to continuously program, reprogram and change the chip,” said Ramanathan.

Romanatan students
Michael Park (left) and Lee Wang, Ph.D., Purdue. students test and analyze a chip designed to simulate human brain learning strategies. (Photo by Purdue University / Rebecca McElho) Download image

Before building a brain in the form of a chip

The equipment is a small rectangular device made of a material called perovskite nickel, which is very sensitive to hydrogen. The application of electrical pulses of different voltages allows the device to change the concentration of hydrogen ions in a matter of nanoseconds, creating a state that, the researchers found, can be compared with the corresponding functions in the brain.

If the device has more hydrogen in the center, for example, it can act as a neuron, a single nerve cell. With less hydrogen at this location, the device serves as a synapse, a connection between neurons that the brain uses to store memory in complex neural circuits.

Using experimental data modeling, Purdue’s team from the University of Santa Clara and the University of Portland showed that the internal physics of this device creates a dynamic structure for an artificial neural network that can more efficiently recognize electrocardiogram patterns and numbers than static ones. networks. This neural network uses “reservoir computing” that explains how different parts of the brain communicate and transmit information.

Researchers from the University of Pennsylvania also demonstrated in this study that as new challenges are presented, a dynamic network can “choose” which circuits are best suited to address these challenges.

Because the team was able to build the device using standard semiconductor-compatible manufacturing methods and work with the device at room temperature, Ramanathan believes the technique can be easily adopted by the semiconductor industry.

“We have demonstrated that this device is very reliable,” said Michael Park, Ph.D. student materials. “After programming the device for a million cycles, reconfiguration of all functions becomes amazingly reproducible.”

Researchers are working to demonstrate these concepts on large-scale test chips that will be used to create a brain-inspired computer.

The experiments in Purdue were conducted at FLEX Lab and Nanotechnology Center Tag with Purdue Discovery Park. Teams from the Argonne National Laboratory, the University of Illinois at Chicago, the Brookhaven National Laboratory and the University of Georgia measured the properties of the device.

The study was supported by the U.S. Department of Energy’s Office of Science, the Air Force’s Office of Research, and the National Science Foundation.

About Purdue University

Purdue University is a leading public research institution that develops practical solutions to today’s most complex problems. In each of the past four years as one of the 10 most innovative universities in the United States according to US News & World Report, Purdue provides research that is changing the world and discovering beyond that world. Committed to practical and online, real-world learning, Purdue offers a transformational education for all. Striving for affordability and affordability, Purdue froze tuition fees and most fees at the 2012-13 level, allowing more students than ever to graduate without debt. See how Purdue never stops in his quest for the next giant leap at https://purdue.edu/.

Writer, media representative: Kayla Wiles, 765-494-2432, wiles5@purdue.edu

Source: Shriram Ramanathan, shriram@purdue.edu


ABSTRACT

Reconfigured electronics made of nickel from perovskite for artificial intelligence

Hai-Tian Zhang, Tae Joon Park, ANM Nafiul Islam, Dat SJ Tran, Sukri Manna, Qi Wang, Sandip Mondal, Haoming Yu, Suvo Banik, Shaabo Chen, Hua Zhou, Sampat Gamag, Sayantan Mahapatra, Imei Zhu, Johannes Abate Nan Jiang, Sub-Ramanan Sankaranarayan, Abronil Sengupta, Christoph Tausher and Shriram Ramanathan

DOI: 10.1126 / science.abj7943

Reconfigured devices offer the ability to program electronic circuits on demand. In this paper, we have demonstrated the creation of on-demand artificial neurons, synapses, and memory capacitors in NdNiO perovskite3 devices that can simply be reconfigured for a specific purpose using disposable electrical pulses. The sensitivity of the electronic properties of perovskite nickel to the local distribution of hydrogen ions allowed to obtain these results. Using experimental data from our memory capacitors, the reservoir computing platform simulation results showed excellent performance for tasks such as digit recognition and classification of heart rate activity on an electrocardiogram. Using our reconfigured artificial neurons and synapses, the simulated dynamic networks outperformed static networks for gradual learning scenarios. The ability to create the building blocks of brain-inspired computers, on demand, opens up new directions in adaptive networks.

Leave a Comment