UW – MSR Machine Learning workshop 2015 – Session 4

14:15Commonsense, Vision and Language – Larry Zitnick The recent significant advances in computer vision, natural language processing and other related areas has led to a renewed interest in artificial intelligence applications spanning multiple domains. In this talk, I explore the relation between computer vision, language and commonsense reasoning through the application of image caption generation. Specifically, I describe new approaches for generating captions using recurrent neural networks, and the use of abstract scenes for gathering commonsense and semantic knowledge. The limitations of current approaches and the challenges that lie ahead are both emphasized.

14:40Machine Learning in Genomics and Proteomics – William S Noble High throughput technologies, including next generation DNA sequencing and tandem mass spectrometry, are producing big biological data sets that cannot be manually interpreted. In this talk, I will describe some of the analysis challenges associated with such data sets, as well as machine learning methods to address these challenges. For example, a grand challenge in genomics is to assign biological meaning (functional annotation) to positions in the human genome. We have developed an unsupervised dynamic Bayesian network methodology to assign such annotations to heterogeneous collections of cell type-specific genomic data. Recently, we have extended this method using graph-based posterior regularization to take into account a new type of genomic data that quantifies the 3D structure of the DNA in the nucleus. In proteomics, we have developed a series of machine learning methods to map from large collections of observed mass spectra to associated proteins and protein subsequences.

15:05Spotlight: Streaming Variational Inference for Bayesian Nonparametric Mixture Models – Alex Tank In theory, Bayesian nonparametric (BNP) models are well suited to streaming data scenarios due to their ability to adapt model complexity with the observed data. Unfortunately, such benefits have not been fully realized in practice; existing inference algorithms are either not applicable to streaming applications or not extensible to BNP models. For the special case of Dirichlet processes, streaming inference has been considered. However, there is growing interest in more flexible BNP models building on the class of normalized random measures (NRMs). We work within this general framework and present a streaming variational inference algorithm for NRM mixture models. Our algorithm is based on assumed density filtering (ADF), leading straightforwardly to expectation propagation (EP) for large-scale batch inference as well. We demonstrate the efficacy of the algorithm on clustering documents in large, streaming text corpora.

15:10Spotlight: Decomposable Norm Minimization with Proximal-Gradient Homotopy Algorithm – Reza Eghbali We study the convergence rate of proximal-gradient homotopy algorithm for norm-regularized linear least squares problems. Homotopy algorithm reduces regularization parameter in a series of steps, and uses proximal-gradient algorithm to solve the problem at each step. Proximal-gradient algorithm has a linear rate of convergence given that the objective function is strongly convex and the gradient of the smooth component of the objective function is Lipschitz continuous. In general, the objective function in this type of problems is not strongly convex, especially when the problem is high-dimensional. We will show that if the linear sampling matrix and the regularizing norm satisfy certain assumptions, proximal-gradient homotopy algorithm converges with a linear rate even though the objective function is not strongly convex. Our result generalizes results on the linear convergence of homotopy algorithm for l1-regularized least squares problems.

Speaker Details

William Stafford Noble (formerly William Noble Grundy) received the Ph.D. in computer science and cognitive science from UC San Diego in 1998. After a one-year postdoc with David Haussler at UC Santa Cruz, he became an Assistant Professor in the Department of Computer Science at Columbia University. In 2002, he joined the faculty of the Department of Genome Sciences at the University of Washington. His research group develops and applies statistical and machine learning techniques for modeling and understanding biological processes at the molecular level. Noble is the recipient of an NSF CAREER award and is a Sloan Research Fellow.

Date:
Speakers:
Larry Zitnick, William S. Noble, and Reza Eghbali
Affiliation:
Microsoft Research, University of Washington
    • Portrait of Jeff Running

      Jeff Running

    • Portrait of Larry Zitnick

      Larry Zitnick

      Principal Researcher