Scientist uses nanomagnets to build brain-like AI, cut power use, boost efficiency
When placed in networks, nanomagnets feel each other much like neurons in the brain.
Benjamin Jungfleisch, an associate professor at the University of Delaware in the US, is intrigued by the human brain’s information processing and retention capabilities. He wants to replicate this for computers but without using electronics and instead relying on nanomagnets. The researcher is confident his work can help lead to a future of energy-efficient artificial intelligence (AI) systems. **
#technology #science #delaware #ai #brain #technology #energy
The race to build AI systems after the popularity of ChatGPT a few years ago has brought multiple concerns about its use. Primarily among them is the energy consumption of these systems during both the training and inference stages.
As AI adoption increases, a larger number of data centers will be needed to power them, leading to a surge in energy consumption in the coming years. If an energy-efficient approach to processing this data isn’t found, the energy demand will harm the planet.
What are magnons?
Just like neurons are the basic unit of the human brain where information processing occurs, magnons are fundamental quantum excitations that make up the spin waves or magnetic waves in a magnetic system.
Magnon connections are like synapses between neurons, which can send signals along specific circuit routes. Jungfleisch works with nanomagnetic arrays and intends to use them as the neuronal network in the brain to process and transmit data.
Advantages of using nanomagnets
The interacting nanomagnets present a distinct advantage. Unlike the traditional computing setup that uses separate memory and processor units, nanomagnets double up to perform both tasks, making it highly efficient. This can help make energy-efficient processes and then build them to perform tasks such as making a chatbot or even creating images.
The operation of such a system is independent of electrons since magnetic excitations are used to store and process data. Interestingly, nanomagnets can keep a history of their states, and they can also be trained.
Jungfliesch and his collaborators are now working on this aspect. Training cycles currently take two to three hours to complete. In the future, though, the team expects the process to be completed in a few minutes.
In a recent publication, the researchers described a three-dimensional nanomagnetic structure that can be easily fabricated and read with existing techniques.
“You get many more states in your system and a much smaller footprint,” added Jungfleisch in the press release. “Storing more information in these networks is easier since you have more space available.”
Article