Artificial neural networks are key components of AI technology, and researchers are always looking for ways to make them more efficient. One area where improvement is needed is in the processing of time-dependent information, such as audio and video data. A recent study led by the University of Michigan has made a significant breakthrough in this area by developing a memristor with a ‘relaxation time’ that can be tuned to mimic the timekeeping mechanism in biological neural networks.
Memristors are electrical components that store information in their electrical resistance, and they have the potential to reduce AI’s energy needs by a factor of 90 compared to current graphical processing units. This is significant given that AI is projected to account for about half a percent of the world’s total electricity consumption in 2027, with that number expected to increase as more companies adopt AI technologies. The goal of this research is to improve the energy efficiency of AI chips by integrating memristors that can mimic the functionality of both artificial and biological neural networks.
The traditional approach to increasing the efficiency of AI networks is to simply increase the size of the network, but this is not very effective. GPUs, which are commonly used to power AI algorithms, require sequential loading of the entire network and all its interactions from external memory, which consumes both time and energy. Memristors offer a more efficient solution as they mimic key aspects of the way that both artificial and biological neural networks function without the need for external memory. By incorporating memristors into AI chips, energy efficiency can be improved sixfold over current state-of-the-art materials without varying time constants.
In biological neural networks, timekeeping is achieved through relaxation. Neurons have different relaxation times, which helps to sequence events and process time-dependent information. In contrast, memristors operate by changing how much of an electrical signal gets through, with relaxation causing the resistance to rise over time. The research team at the University of Michigan has developed a new material system that can provide different relaxation times, allowing memristor networks to mimic the timekeeping mechanism of biological neural networks.
The team built the memristors on a superconductor YBCO, which served as a guide for the organization of the other materials in the memristor. By changing the ratios of these materials, the team achieved time constants ranging from 159 to 278 nanoseconds. The memristor network they built was able to recognize the sounds of numbers zero to nine before the audio input was complete. While the materials used in this study were made through an energy-intensive process, the researchers believe that a simpler process could be used for mass manufacturing, making these materials scalable and affordable.
Overall, this research represents a significant step towards improving the energy efficiency of AI systems by integrating memristors that can mimic the timekeeping mechanism of biological neural networks. With further development and optimization, these memristors could help reduce the energy consumption of AI technology significantly. The researchers are optimistic about the scalability and affordability of the materials used in this study, opening up possibilities for widespread adoption of this technology in the future.
