UT San Antonio Details AI Chip Designed to Learn Without Erasing Old Knowledge

UT San Antonio has detailed how its experimental Genesis AI chip is designed to keep learning without overwriting earlier knowledge. In a September 30 account, the university explains a design that protects important connections while leaving others open to change. The goal is learning inside devices that cannot depend on the cloud—not a commercial chip release. Genesis remains in testing.

Protecting memory without freezing learning

Developed by the university’s MATRIX AI Consortium, Genesis is a neuromorphic accelerator: specialized computing hardware inspired by the brain. Its target is catastrophic forgetting, the problem in which learning a new task can overwrite knowledge gained from an earlier one. The researchers want the device to accumulate knowledge throughout its working life, rather than sacrifice an old skill to acquire a new one.

The approach does not make every connection equally resistant to change. Each processing element tracks its activity and history, including how often it fires and how much it has contributed. Connections judged important resist overwriting; new learning is directed toward those that remain flexible. This principle is called metaplasticity—the regulation of how readily a connection can change.

That balance has consequences beyond keeping a useful memory. In KSAT’s earlier coverage, consortium director Dhireesha Kudithipudi said changing environments require systems to learn while still performing well. Forgetting previous knowledge, she warned, could make decisions dangerous or compromised in some settings. The September 30 explanation describes how the team is trying to address that concern at the hardware level.

Less activity, less data movement

Genesis’s efficiency argument has two parts. First, it uses spiking neural networks, which process information through pulses resembling the signals of biological neurons. When there is no task to perform, the system rests. The university contrasts this approach with the standard artificial neural networks used in large language models, presenting reduced activity as one route to lower power use.

Second, the design reduces the travel between computation and memory. Researcher Fatima Tuz Zohora told KSAT that Genesis 2.0 uses memristors, compact memory devices arranged in a crossbar structure. They can perform computation in the same physical material that stores memory. Her explanation connects the hardware choice directly to reduced data movement and energy consumption during continual learning.

The intended setting is the edge: devices doing their own computing where a data center or reliable cloud connection is unavailable. The university says Genesis runs on milliwatts and identifies implantable medical devices, field-deployed drones and wearable sensors as potential applications. Those are intended uses, not deployments. Researcher Vedant Karia also described it as an additional low-power accelerator for equipment such as healthcare devices and robots.

Fabricated silicon, unfinished deployment

Genesis is more than a proposed architecture. The chips are fabricated through a partnership with SUNY Albany using IBM’s 65-nanometer technology. The team worked through two earlier prototypes over five years before reaching the current architecture. Alongside the hardware, researchers developed learning algorithms and MetaplasticNet, a brain-inspired neural-network architecture, bringing those components together to support low-power learning.

The work is supported by a multimillion-dollar, five-year Air Force Research Laboratory grant. Kudithipudi told KSAT that design and fabrication cost about $2 million. The next phase, according to the university, is to scale the learning mechanisms, prepare them for real-world deployment and integrate them with new hardware and software. Karia’s assessment remains the practical boundary: additional development is needed before the chip is fully deployable.