SAN MARCOS – Texas State University researchers Ziliang Zong, associate professor, and Yan Yan, assistant professor in the Department of Computer Science, have received a $500,000 grant from the National Science Foundation to develop more energy-efficient and compact artificial intelligence (AI) technology.
The three-year grant was awarded by the Division of Computer and Network Systems and will fund Zong and Yan’s project, “Interpretable Multi-Modal Neural Network Pruning for Edge Devices.”
AI systems, which use machine learning and deep logic techniques, have improved rapidly in recent years. Many AI tasks, such as image recognition and audio processing, have improved to the point where they are comparable to the capabilities of humans. The difficulty comes from the fact that the software driving these sophisticated AI networks is very large and consumes tremendous amounts of energy, making it impractical for use in common applications.
“The CO2 emissions released by training a single AI model of Natural Language Processing using the architectural search methodology was equal to that of five cars during their lifetimes,” said Zong. “It has become unsustainable and environmentally unfriendly.
“The other downside, of course, is the size of the AI models. We want to use more AI on smartphones, but the model’s too large,” he said. “The AI takes up so much memory, which makes it hard to fit into phones or other smart devices. Not only do we have to care about the AI’s accuracy, but we also have to care about how small, how efficient the network is. If we cannot do that, we cannot apply AI everywhere.”
Yan and Zong are working to solve the problem from two directions. They are developing pruning and compression technology that won’t impact the accuracy of the AI. If successful, this will result in AI models that can be loaded into smartphones and other devices collectively known as the Internet of Things. They are also tackling the energy consumption problem. Because it’s a computation-intensive process, AI consumes far more power than regular apps, even games. This poses a significant problem for devices that depend on extended battery life.
Manufacturing and other robotics-dependent industry also stands to benefit. When AI models are very large, processing times dramatically increase. At their most efficient, industrial robots complete tasks at the millisecond level. Complex AI processing can slow robotic decision-making down to a half second or even longer, which is unacceptable on assembly lines. Faster, smaller, more efficient AI would solve this issue.
“Current AI is not like human intelligence,” Yan explained. Traditionally, AI is designed to focus on a task that roughly corresponds to human senses, that is, image recognition is akin to human vision, speech recognition is akin to human hearing, etc. Rather than dedicate a specific AI to each set of sensory inputs, the Texas State project proposes a multi-modal approach to all the senses, in which a single AI is flexible enough to interpret the different channels of incoming data.
“Our goal is a more general type of AI model that can synthesize all this information together,” Yan said. “Our hope would be that this increases the accuracy and gives us a better understanding.”
The researchers already have three doctoral students assisting in their research, and as the project matures, they will involve talented undergraduate students in the coming year.
Yan and Zong have already had a paper based on preliminary results accepted, which they will present during the IEEE Winter Conference on Computer Vision in March 2020.
“We always want to go to a conference first with the latest, newest ideas. Then we’ll take our extended version to a journal for publication,” Zong said. “We’re excited to see our preliminary results recognized by the community as valuable work.”