Strengthening the field to pull useful information from increasingly complex, increasingly big data

Machine learning is already changing everyday life. Email providers use machine learning to choke off the flood of spam heading for our inboxes, banks use it to spot credit card fraud and streaming music services us it to select songs we like. In all these cases, these organizations apply algorithms that analyze data for specific characteristics, identify significant patterns and use these patterns to make predictions.

But there is much more that machine learning could do. “The past decade has seen a surge of research on how best to deploy machine learning to extract useful information from large quantities of complex data,” said Quanquan Gu, an assistant professor of computer science who holds a joint appointment in the Department of Systems and Information Engineering. “We are seeing it being applied in fields as diverse as cyber-physical systems and biomedical data science.”

The challenge these researchers face is that as the quantity of data being analyzed becomes exponentially larger, more complex and multidimensional, traditional techniques of mining the data and optimizing the resulting decisions cannot keep pace. Because his research straddles the computer science discipline of machine learning and the systems engineering discipline of optimization, Gu is ideally placed to address this challenge.

Researchers have applied machine learning methods based on nonconvex optimization with significant empirical success in cases of massive amounts of high-dimensional data. The National Science Foundation recently awarded Gu a CAREER Award to develop a new generation of nonconvex statistical optimization algorithms and to provide a theoretical basis for them. “The empirical success of nonconvex optimization cannot be adequately interpreted by existing machine learning and optimization theory,” Gu said. My goal is to provide a systematic way to design nonconvex high-dimensional machine learning methods with strong theoretical guarantees. I want to bridge the gap between theory and practice.”

As Gu points out, nonconvex statistical optimization can be applied to a variety of machine learning approaches. These include deep learning, which uses a multilayered, interconnected neural network that mimics the way the brain processes information, and high-dimensional graphical modeling, which uses a graph to express the structure of dependence between random variables.

The proving ground for Gu’s algorithms are real-world applications — and this requires collaboration with experts in other disciplines. Gu has partnered with faculty members in the psychology department to analyze three-dimensional, functional MRI images. He is using neural networks to correlate brain activity with social interactions. “This is a high-dimensional challenge,” he says. There can be as many as 60,000 voxels (three-dimensional pixels) in a single image.

Gu is also working with experts in computational cancer genomics to link gene regulation to subtypes of cancer. In this case, he is using high-dimensional graphical modeling. “If we are able to identify the unique gene regulation relationships for specific cancers, we could diagnose the disease more accurately and devise a treatment targeted at biomarkers associated with a given phenotype,” he said.

Although Gu’s core academic interest is to develop the theoretical foundation for algorithms for machine learning and therefore strengthen the field, his desire to better society and help people is behind his choice of applications.

“The work I’m doing has a direct impact on our ability to realize the promise of precision medicine and provide personalized healthcare,” he said. “From the point of view of machine learning, human health is a very hard problem.”