Learning through Fitting: Encoding Priors in Implicit Neural Representations for Visual Inference

Abstract: Gridded pixel and voxel representations form the backbone of visual computing, but they struggle to scale efficiently to large, high-dimensional data, such as volumetric medical scans and complex scientific simulations. Consequently, alternative non-gridded models such as implicit neural representations (INRs) have gained significant traction over the past five years. However, INRs have largely been confined to representing individual signals, rather than acting as useful data types for encoding patterns over a distribution of signals for downstream inference tasks.

In this talk, I will present our recent work on injecting data-driven priors into INRs. By initializing INR parameters over a cohort of signals, we show that the resulting parameters are useful for various downstream applications such as rapid amortized signal fitting, inverse imaging and even semantic segmentation.

Bio: Guha Balakrishnan is an assistant professor in the Electrical and Computer Engineering department at Rice University. His scientific contributions have been recognized with several honors, including the National Science Foundation CAREER Award and the Medical Image Computing and Computer Assisted Intervention Society Best Paper Award. Before joining Rice, he completed his Ph.D. at the Massachusetts Institute of Technology’s Computer Science and Artificial Intelligence Laboratory and earned his undergraduate degrees in Computer Science and Computer Engineering from the University of Michigan, Ann Arbor.

See upcoming and previous presentations at: CS Seminar Series.