The Ken Kennedy Institute's April Member of the Month, Dr. Guha Balakrishnan, is an Assistant Professor of Electrical and Computer Engineering at Rice University. Before joining Rice, Dr. Balakrishnan received his BS (2011) degrees in Computer Science Engineering and Computer Engineering from the University of Michigan, and his MS (2013) and PhD (2018) degrees in EECS from MIT. He was also a postdoctoral researcher at MIT from 2018-2020, and an applied scientist at Amazon Web Services (AWS) from 2020-2021 working on fairness of AI systems.
Dr. Balakrishnan is interested in the theory, practical design, and downstream applications of generative models for complex visual data. He is particularly excited by their applications to promote fairness and accountability of vision systems. Guha has also worked on a broad range of medical applications, including developing ML algorithms for medical image registration and remote vital signs measurement from video.
What is your favorite book?
All Quiet on the Western Front, Shogun, Pride and Prejudice.
How do you explain your research in one sentence?
I develop machine learning models that can generate new image data, particularly for applications in medical imaging and measuring/mitigating biases of computer vision models.
What is your favorite aspect of your research?
I enjoy the fact that applications in computer vision and graphics can have large, immediate impacts. This is in part because image data are so widespread and important in our society.
What challenges do you see in your research that you didn't expect?
Modern computer vision algorithms typically contain huge numbers of parameters that need to be trained over millions of data points. As a result, industrial research labs are better equipped to achieve state-of-the-art results on many vision tasks due to their resources. This poses a challenge to academic researchers: we must find important research directions where we may make important contributions without needing enormous resources.
How do you see computation and data advancing in the future?
Machine learning models now achieve excellent performance on typical supervised learning tasks. As a result, we are starting to see more nuanced evaluations and designs of these models based on concepts like explainability, generalization to out-of-distribution samples, and causality. I think this will only grow in importance in the future.
