A challenge that has been holding back machine learning and has the potential to revolutionize technologies such as voice recognition , image processing and autonomous driving could have been unlocked by the discovery of a new method for making non-volatile computer memory.
In the peer-reviewed journal Advanced Materials, a team from Sandia National Laboratories, working with collaborators from the University of Michigan, published a paper outlining a new approach that will combine computer chips that power machine-learning applications with more processing power through the use of a common material found in house paint in an analog memory system that allows highly energy-efficient processing power.
“One of the most widely produced materials is titanium oxide. It contains titanium oxide in any paint you purchase. It’s affordable and non-toxic,” explains Alec Talin, Sandia materials scientist. “It’s an oxide, there’s oxygen already. But you create what are called oxygen vacancies if you take a few out. It turns out that you make this substance electrically conductive when you create oxygen vacancies.”
Those vacancies in oxygen can now store electrical data, giving more computing power to almost any system. By heating a computer chip with a titanium oxide coating above 302 degrees Fahrenheit (150 degrees Celsius), Talin and his team generate oxygen vacancies, using electrochemistry to remove some of the oxygen molecules from the material and generate vacancies.
“It stores any data for which you program it when it cools off,” Talin said.
A boost to machine learning for energy efficiency
Computers usually function right now by storing information in one place and processing the information in another place. That means computers have to continuously move data, wasting energy and computing power from one location to the next.
Yiyang Li, the lead author of the paper, is a former Truman Fellow at Sandia and now an assistant professor of materials science at Michigan University. He clarified how their method has the ability to change the way computers operate fully.
“What we did was do the manufacturing and storage at the same location,” Li said. “What’s new is that, in a consistent and repeatable way, we were able to do it.”
He and Talin both see the use of oxygen vacancies as a way of helping machine learning resolve a major challenge that is currently holding it back — power consumption.
If we try to do machine learning, it takes a lot of energy because you transfer it back and forth and power consumption is one of the obstacles to the realization of machine learning, “Li said.” “When you have autonomous cars, it takes a huge amount of energy to process all the inputs to make driving decisions. If we can develop an alternate material for computer chips, they will be able to more effectively process information, save energy and process a lot more data.”
Everyday research affects research
In the success of everyday devices, Talin sees the promise.
“Think about your mobile phone,” he said. “You need to be linked to a network that sends the command to a central hub of computers that listen to your voice if you want to give it a voice command, and then send a signal back to tell your phone what to do. Voice recognition and other functions happen right on your phone through this process.”
Talin said the team is working on optimizing many procedures and processes.
Talin said the team is working on optimizing multiple processes and larger-scale testing of the system. The project is sponsored by the Laboratory Guided Research and Development program of Sandia.
Source of Story: DOE / Sandia National Laboratories supplied the components. Note: For style and length, material can be edited.
No Comments
Leave a comment Cancel