Uploaded July 2026 | Updated November 2026, 1 day ago
Speaker: Tim Holy (Washington University in St. Louis)
Abstract: Planar and volumetric calcium imaging data sets are often analyzed via nonnegative matrix factorization (NMF), splitting the observed imaging data into components which one hopes corresponds to individual neurons. This problem faces several challenges, including the existence of multiple local minima of varying quality and challenges in applying NMF to data sets that range up to terabyte scale. In this talk I will describe a new class of tools for NMF, variable rank methods, that result more reliably in high-quality solutions without substantial increases in solver time. I will present preliminary results for these techniques on calcium imaging and argue that methods tailored for shot-noise statistics represent a next major frontier for factorization-based analysis of calcium imaging data. Finally, I will talk about computational optimizations, including a new representation of sparse localized components called a TileTree, that scale performance to terabyte images on ordinary workstations.
Speaker: Tim Holy (Washington University in St. Louis)
Abstract: Planar and volumetric calcium imaging data sets are often analyzed via nonnegative matrix factorization (NMF), splitting the observed imaging data into components which one hopes corresponds to individual neurons. This problem faces several challenges, including the existence of multiple local minima of varying quality and challenges in applying NMF to data sets that range up to terabyte scale. In this talk I will describe a new class of tools for NMF, variable rank methods, that result more reliably in high-quality solutions without substantial increases in solver time. I will present preliminary results for these techniques on calcium imaging and argue that methods tailored for shot-noise statistics represent a next major frontier for factorization-based analysis of calcium imaging data. Finally, I will talk about computational optimizations, including a new representation of sparse localized components called a TileTree, that scale performance to terabyte images on ordinary workstations.










