SAN ANTONIO – With all the things artificial intelligence or AI technology can do, there is one thing it cannot do: remember.
“It’s not able to retain knowledge that it has learned in the past,” Dhireesha Kudithipudi, Ph.D, said.
Kudithipudi is the director of UT San Antonio’s Matrix AI Consortium and the Neuromorphic Artificial Intelligence Laboratory (NUAI) Lab in the College of AI, Cyber and Computing.
She said when AI systems in use are programmed to learn new tasks, they can begin to override what it has previously learned. This phenomenon is known as “catastrophic forgetting.”
Kudithipudi said it is critical that AI systems be able to recall previous knowledge it learned and gathered because the technology is being used more and more in critical environments like healthcare and exploration.
“The environment is constantly changing, and you need to be able to learn in that changing environment and still perform well,” she said. “But what if your chip has forgotten what it has learned in the past, the decisions it’s going to make are going to be dangerous in some scenarios and compromised in other scenarios.”
For a decade, Kudithipudi has sought to come up with a solution, and within the last five years she and the members of her research and development team believe they may have it.
Their solution is the Genesis 2.0 AI chip, designed with a continuous learning software or algorithm so the chip itself can keep learning throughout its lifetime.
Its design was inspired by how the human brain forms connections.
Nicholas Soures, Ph.D, is a member of the consortium who worked on the Genesis chip. Soures said working alongside neuroscientists on the team, they were able to look and replicate some features of the brain to build the algorithm.
“We started looking at spiking neural networks ... and features such as metaplasticity and others, looked like a natural fit for a concept of a continued learning system in a platform that was already adaptable to energy efficient hardware,” Soures said.
Fatima Tuz Zohora, Ph.D, was focused on finding the right device that could both run the continuous learning algorithm and also store what the chip learned.
“We use memristors, which are emerging memory devices, to perform continual learning ... These devices are really compact,” Tuz Zohora said. “They are nanoscale devices and arranged in a crossbar setting or in a crossover structure (and) can perform computation in the same substrate where the memory is stored. So this reduces data movement during continual learning and reduces energy consumption.”
Vedant Karia, Ph.D, is also on the team for the Genesis 2.0 chip. Karia said the chip can be connected to any device and, if desired, it could be used as an additional accelerator at low power.
“All the high-energy, high-power computations that traditional systems do can be offloaded to this chip and can be done at a very low power. That was the main idea of this chip while developing,” Karia said. “So it can be connected to any devices in, (for example) you take any healthcare devices or any robotics, you can connect them directly to the processor on them and then you are all set to go.”
Karia said the chip is not ready go mainstream just yet, explaining that few more steps need to be completed before it is completely deployable for real-world application.
Kudithipudi said the progress of the Genesis 2.0 chip as it currently is took about five years, with the design and fabrication costing around $2 million.
Research and development of the chip has been funded through a grant by the Air Force Research Laboratory (AFLR).
Team members are excited and consider the Genesis 2.0 chip a major win in the evolution of true AI.
Read also:
