July 2019 - talks - lists.cs.princeton.edu

Brian Bullins will present his FPO, "Efficient Higher-Order Optimization for Machine Learning" on Tuesday, 7/23/2019 at 2pm in CS 402
by Nicki Mahler 22 Jul '19

22 Jul '19

Brian Bullins will present his FPO, "Efficient Higher-Order Optimization for Machine Learning" on Tuesday, 7/23/2019 at 2pm in CS 402. The members of his committee are: Elad Hazan (Adviser), Examiners: Mark Braverman and Sanjeev Arora; Readers: Yoram Singer and Yuxin Chen (EE). A copy of his thesis, is available upon request. Everyone is invited to attend his talk. The talk abstract follows below: In recent years, stochastic gradient descent (SGD) has taken center stage for training large-scale models in machine learning. Although some higher-order methods have improved iteration complexity in theory, the per-iteration costs render them unusable when faced with millions of parameters and training examples. In this thesis, I will present several works which enable higher-order optimization to be as scalable as first-order methods. The first method is a stochastic second-order algorithm for convex optimization, called LiSSA, which uses Hessian information to construct an unbiased Newton step in time linear in the problem dimension. To bypass the typical eciency barriers for second-order methods, we harness the ERM structure in standard machine learning tasks. While convex problems allow for global convergence, recent state-of-the-art models, such as deep neural networks, highlight the importance of developing a better understanding of non-convex guarantees. In order to handle this challenging setting, I will present FastCubic, a Hessian-based method which provably converges to first-order critical points faster than gradient descent, while additionally converging to second-order critical points. Finally, I will establish how to leverage even higher-order derivative information by means of the FastQuartic algorithm, which achieves faster convergence for both smooth and non-smooth convex optimization problems by combining ecient tensor methods with near-optimal higher-order acceleration.

1 1

Ruth Dannenfelser will present her Pre FPO "From gene expression data to biomedical abstracts: data-driven approaches and systems to interrogate complex disease" on Tuesday, July 16, 2019 at 9am in CS 402.
by Nicki Mahler 11 Jul '19

11 Jul '19

Ruth Dannenfelser will present her Pre FPO "From gene expression data to biomedical abstracts: data-driven approaches and systems to interrogate complex disease" on Tuesday, July 16, 2019 at 9am in CS 402. The members of her committee are: Olga Troyanskaya (adviser), Mona Singh, Kai Li, Yibin Kang (Mol Bio), Wendell Lim (UCSF). Everyone is invited to attend her talk. The talk title and abstract follow below: Title: From gene expression data to biomedical abstracts: data-driven approaches and systems to interrogate complex disease Large-scale genomic studies now give more predictive power than ever, allowing us to profile the composition of tissues, study cellular functions, and understand organismal traits at an unprecedented level of detail. This is particularly important for studying heterogeneous diseases, such as cancer, where small patient-specific differences play critical roles in disease development and progression. Here we build off the wealth of publicly available data to examine the interplay between cancer and the immune system and then develop two query-based visualization systems that enable interactive data exploration for the biomedical community at large. The first part of my talk will present two perspectives on cancer and the immune system, starting with a semi-supervised approach for immune cell type quantification. Using derived immune markers we examine lymphocyte infiltration in breast cancer and find that estrogen receptor activity and genomic complexity are the key factors driving variation in lymphocytic infiltrate across individuals. In a broader scope, we leverage public gene expression data to further the development of targeted immunotherapeutics for solid tumors. Engineered T cell therapies have shown great promise for hematological cancers but have only found limited success in targeting solid tumors due to off target effects. Working closely with experimental collaborators we are developing a method to prioritize safer pairs of antigen targets that will help engineered T cells hone in on specific tumor targets minimizing damage to normal tissues. The second half will cover how we can extract unbiased signals from large collections of biomedical data in the form of scientific abstracts and large repositories of transcriptomics data. First, we show how we can obtain informative tissue-disease-gene relationships from abstracts and integrate them into a web system that presents different snapshots of curated interactions and adds tissue and disease annotations to gene lists from experimental assays (e.g. GWAS, differentially expressed genes, drug screens, etc). Secondly, we extend SEEK, a gene expression search engine that simultaneously returns coexpressed genes and relevant datasets where the genes are likely coregulated. Our extension enables significant performance improvements, expands the search space across the major model organisms, and provides a new cross-organism exploration interface to help facilitate translational research.

1 0

Maciej Halber will present his FPO, "RGBD pipeline for Indoor Scene Reconstruction and Understanding" on Monday, July 15th, 2019 at 1pm in CS 402.
by Nicki Mahler 08 Jul '19

08 Jul '19

Maciej Halber will present his FPO,"RGBD pipeline for Indoor Scene Reconstruction and Understanding" on Monday, July 15th, 2019 at 1pm in CS 402. The members of his committee are Thomas Funkhouser (adviser); Readers: Szymon Rusinkiewicz and Matthias Niessner (The Technical University of Munich); Examiners: Thomas Funkhouser, Olga Russakovsky, and Adam Finkelstein. A copy of his thesis is available upon request. Everyone is invited to attend his talk. The talk abstract follows below: In this work, we consider the problem of reconstructing a 3D model from a sequence of color and depth frames. Generating such a model has many important applications, ranging from the entertainment industry to real estate. However, transforming the RGBD frames into high-quality 3D models is a challenging problem, especially if additional semantic information is required. In this document, we introduce three projects, which implement various stages of a robust RGBD processing pipeline. First, we consider the challenges arising during the RGBD data capture process. While the depth cameras are providing dense, per-pixel depth measurements, there is a non-trivial error associated with the resulting data. We discuss the depth generation problem and propose an error reduction technique based on estimating an image-space undistortion field. We describe the capture process of the data required for the generation of such an undistortion field. We showcase how correcting the depth measurements improves the reconstruction quality. Second, we address the problem of registering RGBD frames over a long video sequence into a globally consistent 3D model. We propose a “fine-to-coarse” global registration algorithm that leverages robust registrations at finer scales to seed detection and enforcement of geometrical constraints, modeled as planar structures, at coarser scales. To test global registration algorithms, we provide a benchmark with 10,401 manually-clicked point correspondences in 25 scenes from the SUN3D dataset. We find that our fine-to-coarse algorithm registers long RGBD sequences better than previous methods. Last, we show how repeated scans of the same space can be used to establish associations between the different observations. Specifically, we consider a situation where 3D scans are acquired repeatedly at sparse time intervals. We develop an algorithm that analyzes these “rescans” and builds a temporal model of a scene with semantic instance information. The proposed algorithm operates inductively by using a temporal model resulting from past observations to infer instance segmentation of a new scan. The temporal model is continuously updated to reflect the changes that occur in the scene over time, providing object associations across time. The algorithm outperforms alternate approaches based on state-of-the-art networks for semantic instance segmentation.

1 0

Elena Balashova will present her FPO, "Shape Synthesis Using Structure-Aware Reasoning" on Monday, July 8, 2019 at 1pm in CS 105
by Nicki Mahler 08 Jul '19

08 Jul '19

Elena Balashova will present her FPO, "Shape Synthesis Using Structure-Aware Reasoning" on Monday, July 8, 2019 at 1pm in CS 105. The members of her committee are as follows: Thomas Funkhouser (adviser), Readers: Szymon Rusinkiewicz and Vivek Singh (Siemens Healthcare); Examiners: Olga Russakovsky, Adam Finkelstein, and Thomas Funkhouser (adviser). A copy of her thesis is available upon request. Everyone is invited to attend her talk. The talk abstract follows below. Shape synthesis is an important area of computer vision and graphics that concerns creation of new shapes and reconstruction from partial data. Its goal is to learn a model that can generate shapes within an object category suitable for novel shape creation, interpolation, completion, editing, and other geometric modeling applications. Existing tools learn shape properties from large collections of shapes. Although these methods have been very successful at learning how to synthesize the coarse shapes of objects in categories with highly diverse shapes, they have not always produced examples that reconstruct important structural elements of a shape. In this thesis, I describe how structure can be incorporated into the synthesis process, and how it can be used to improve generative models. First, I introduce a template-defined skeleton structure for learning a part-aware generative model in typography, where the shapes have a known structure and can be explained by a small number of templates. Next, I present a scenario of noisy archaeological wall painting (fresco) reconstruction from eroded fragments, where there is no well-defined structure and exponentially many arrangement possibilities in this case, I present a cluster evaluation function that guides the assembly process and encourages selection of good clusters. Finally, I describe a semantic landmark-based structure and how it can be used to improve a generative model of examples with extremely varied topology by means of a geometric shape-structure consistency loss. Through exploration of each type of structure, I show how reasoning with proposed structures helps synthesize more accurate and realistic shapes. I also propose a fully automatic framework for font completion. Finally, I design a genetic algorithm for wall painting reconstruction and propose an iterative outlier detection technique based on the eigenvector method.

1 1