Using the human brain as inspiration, a team of researchers at the University of Texas at San Antonio created Genesis 2.0, a chip that could be an answer to a known artificial intelligence shortcoming called “catastrophic forgetting.”
This is the inability of AI systems to continue learning without losing, or forgetting, older information.
“Our team has specifically taken inspiration from biology,” said Dhireesha Kudithipudi, founding director of UT San Antonio’s MATRIX AI Consortium and the Neuromorphic Artificial Intelligence Laboratory in the College of AI, Cyber and Computing.
“High-energy efficiency and the high degree of plasticity, … those drew our attention to build machines that can learn similar to humans throughout their lifetime,” she added.
Kudithipudi and a team of doctoral students and postdoctoral fellows — Vedant Karia, Fatima Tuz Zohora, Abdullah Zyarah and Nicholas Soures — set out to try to mimic how the human brain operates.
They worked closely with neuroscientists to understand how to create algorithms and hardware to help machines learn without forgetting and use low energy while at it.
But the challenge wasn’t as simple as increasing the storage capacity of an AI system, Kudithipudi said.
Much like the human brain, the goal was to help it determine what information is useful to its learning process so that it can create stronger connections between the information over time and perhaps weaken other connections as it learns more.
“It’s not just purely adding buffer to replay all the past experiences, but rather trying to identify what information is important or significant as new tasks are coming into the model or into the network,” she said.
The outcome is Genesis 2.0, a chip that replicates the human brain learning process accumulating knowledge and tracking what is currently learning and what it has learned before. It also has the ability to save energy by processing information in what they call “spikes” and saving energy when it is at rest.
“Today the Genesis chip operates under 20 milliwatts to a little over 20 milliwatts, that is what you need to power up a single LED or maybe your phone screen,” Kudithipudi said.
The university landed a five-year grant by the Air Force Research Laboratory for this project. Other grants from the National Science Foundation are being used to support extensions of this research, Kudithipudi said, allowing her and her team to continue with research and development of this technology.
Soures, who focused on developing the algorithms, joined the team from New York, where he is part of the Lifelong Learning Machines Program by the Defense Advanced Research Projects Agency, or DARPA.
“I very quickly got into working with different neuroscientists and different groups of researchers to learn about the neuroplasticity in the brain and seeing what was not implemented in the current learning rules that most AI models use,” Soures said.
Aligning the algorithm, or software, and hardware was always top priority, the team said. Throughout the process they had to make tradeoffs to make sure one worked well with the other and led to their ultimate goal of creating a product that’s useful in real-world applications.
That is how the team landed at Genesis 2.0, a chip that can operate longer without a recharge, retain information even when it has no access to offloading data, and continue learning while making connections that can inform its purpose.
The university partnered with State University of New York-Albany to fabricate the chips using IBM’s 65-nanometer technology, a manufacturing process developed to shrink the size of a semiconductor device and operate at faster speeds with lower power consumption.
This chip can be used for small wearable or implantable devices for healthcare purposes, field-deployed drones, wearable sensors, and more. It can also help inform other potential developments.
“In Genesis 2.0, we investigated one type of emerging memory, but there are many different kinds of emerging memory devices, each with their unique strengths,” Tuz Zohora said. “So, one of the things that I am excited about is, how can we use their unique strengths together so that this leads to even more efficiency and we can use every component to its best?”