Nebraska Researcher Builds AI Model Using 15 Parameters
A new architecture inspired by bacterial gene networks reduces AI complexity by a factor of 690, enabling low-power edge computing.
Key points
- The new AI architecture uses 15 parameters in one test, compared to over 10,000 in conventional models.
- Research is funded by a $782,358 National Science Foundation grant for three years of development.
- The system is designed for edge computing, targeting devices with strict energy and size constraints.
Sasitharan Balasubramaniam, an associate professor at the University of Nebraska School of Computing, has secured a $782,358 grant from the National Science Foundation to develop a new type of artificial intelligence. This project moves away from traditional neural networks, which mimic the human brain, and instead draws inspiration from the gene regulatory networks found in bacteria.
Current AI systems typically require significant computational power and energy, limiting their use in small or remote devices. By studying how bacteria process information with minimal energy, Balasubramaniam aims to create an "artificial non-neuronal network" that is both lightweight and highly efficient, according to a report by Nebraska Today.
Bacteria as a computational model
Although bacteria lack brains or nervous systems, they can sense their environment and make survival decisions using interconnected gene control systems. Balasubramaniam notes that these organisms adapt to harsh conditions with remarkable resilience and energy efficiency. The research team plans to analyze these biological mechanisms to build mathematical models that replicate this efficient decision-making process.
Significant reduction in complexity
Preliminary results indicate a dramatic decrease in the number of parameters needed for accurate predictions. In one test case, a conventional neural network required over 10,000 parameters, while the new bacterial-inspired approach achieved lower prediction errors with only 15 parameters. This represents a reduction of roughly 690 times, pointing toward much more compact computing systems.
Enabling low-power edge devices
The ultimate goal is to deploy this architecture on field-programmable gate arrays, which are integrated circuits designed for specific, re-programmable tasks. By reducing energy consumption and network size, the technology could make AI viable for tiny devices and energy-constrained settings where current systems are impractical. The team will benchmark performance against standard metrics to ensure the new framework maintains scalability and functionality.
