Stats, Optimization, & Machine Learning Seminar - Derek Driggs
| Event Description: Seminar meets in DUAN G2B21. Derek Driggs, Department of Applied Mathematics, Â鶹ӰԺ Faster Low-Rank Recovery Low-rank recovery models seek to decompose data sets into their underlying structured components. They have been used to impute missing data points, identify people in busy surveillance video, and recover important information from noisy signals. As many other machine-learning algorithms turn to parallelized computing to broaden their applications, low-rank recovery models are limited by their high communication costs. In this talk, we present two ways to accelerate low-rank recovery. The first method uses splitting and smoothing techniques to transform low-rank models into non-convex programs with provable recovery guarantees. The second method uses a communication-avoiding QR decomposition to accelerate the randomized SVD on parallelized architectures. We will show that the enhanced efficiency due to both of these methods allows low-rank recovery models to find new applications, such as real-time video processing and denoising fMRI brain scans. Ìý |
| Location Information: ÌýÌý() 2000 COLORADO AV Â鶹ӰԺ, CO |
| Contact Information: Name: Ian Cunningham Phone: 303-492-4668 Email: amassist@colorado.edu |