1,116 research outputs found
Polarization and social media: a systematic review and research agenda
No full textThe world of today presents a duality of phenomenal progress and persistent ills. Amid such existential contradictions, public deliberation forms one of the central pillars of a functional and progressive society. Though its relevance remains undoubted, interactions in the public sphere may often give way to misinformation, affect-driven predisposition, and homophily-based interactions, all reminiscent of polarization. While polarization remains a concern worldwide, structural changes, most notably, social media's advent and remarkable progress, have further redefined the meaning, scale, and diffusion of information. Accordingly, a tireless debate rages regarding the valence and strength of social media's influence on polarization. As an incremental means of resolving the complexity, we perform a systematic review of the extant scholarship and identify contingencies and mechanisms of social media's relationship with polarization. Further, we provide a conceptual framework, incorporating these intricacies while emphasizing the need to place this association in a broader frame. Our work contributes to theory by being one of the few reviews linking social media to polarization and providing a synthesis of contingent factors and underlying processes. We guide policy and practice by suggesting a future research framework
The Quest for Mathematical Understanding of Deep Learning (Invited Talk)
No full textDeep learning has transformed Machine Learning and Artificial Intelligence in the past decade. It raises fundamental questions for mathematics and theory of computer science, since it relies upon solving large-scale nonconvex problems via gradient descent and its variants. This talk will be an introduction to mathematical questions raised by deep learning, and some partial understanding obtained in recent years
Overcoming Intractability in Unsupervised Learning (Invited Talk)
No full textUnsupervised learning - i.e., learning with unlabeled data - is increasingly important given today's data deluge. Most natural problems in this domain - e.g. for models such as mixture models, HMMs, graphical models, topic models and sparse coding/dictionary learning, deep learning - are NP-hard. Therefore researchers in practice use either heuristics or convex relaxations with no concrete approximation bounds. Several nonconvex heuristics work well in practice, which is also a mystery.
The talk will describe a sequence of recent results whereby rigorous approaches leading to polynomial running time are possible for several problems in unsupervised learning. The proof of polynomial running time usually relies upon nondegeneracy assumptions on the data and the model parameters, and often also on stochastic properties of the data (average-case analysis). We describe results for topic models, sparse coding, and deep learning. Some of these new algorithms are very efficient and practical - e.g. for topic modeling
Ideas for rent: an overview of markets for technology
No full textThis article surveys some of the recent literature on technology markets, and summarizes its main issues and insights. We structure our analysis in three parts: the supply and demand of technology; the factors that condition the formation and growth of technology markets; industry structure and dynamic issues. In addition, we summarize some of the studies that have tried to document the size and growth of these markets. We find that the literature has focused mainly on the supply of technology, but several other aspects of these markets remain under-studied, including the demand for external technology, the role of uncertainty in technology markets, and the dynamic interaction between industry structure and the market for technology. Understanding these will illuminate whether markets for technology will continue to grow or remained confined to pockets of the economy. Copyright 2010 The Author 2010. Published by Oxford University Press on behalf of Associazione ICC. All rights reserved., Oxford University Press.
Metrics for analytics and visualization of big data with applications to activity recognition
No full textActivity recognition systems detect the hidden actions of an agent from sensor measurements made on the agents' actions and the environmental conditions. For such systems, metrics are important for both performance evaluation and visualization purposes. In this thesis, such metrics are developed and illustrated. For human activity recognition datasets, a reporting structure is described to visualize the metrics in a systematic manner. The other contribution of this thesis is to describe a visualization tool for estimating the orientation (attitude) of a rigid body from streaming motion sensor (accelerometer and gyroscope) data. A feedback particle filter (FPF) is implemented algorithmically to solve the estimation problem.Submission published under a 24 month embargo labeled 'Closed Access', the embargo will last until 2018-05-01The student, Rohan Arora, accepted the attached license on 2016-04-25 at 10:47.The student, Rohan Arora, submitted this Thesis for approval on 2016-04-25 at 10:48.This Thesis was approved for publication on 2016-04-27 at 15:05.DSpace SAF Submission Ingestion Package generated from Vireo submission #9459 on 2016-07-07 at 14:17:57Made available in DSpace on 2016-07-07T21:18:02Z (GMT). No. of bitstreams: 2
ARORA-THESIS-2016.pdf: 2048739 bytes, checksum: f76095ae5ef05e4ce14c6b05ab503f5d (MD5)
LICENSE.txt: 4208 bytes, checksum: e5888a1be6c205bee6e88396c3d3da15 (MD5)
Previous issue date: 2016-04-27Embargo set by: Seth Robbins for item 93308
Lift date: 2018-07-07T21:18:16Z
Reason: Author requested closed access (OA after 2yrs) in Vireo ETD systemLimited Restriction Lifted for Item 93308 on 2018-07-08T09:15:30Z
The missing middle: value capture in the market for startups
No full textWe argue that innovations that involve both upstream (technological) and downstream (commercialization) challenges are disadvantaged in a startup-based innovation system where startups develop inventions, while incumbents acquire startups. We propose an analytical model in which startups are more efficient at solving technological challenges and incumbents are more efficient at solving commercialization challenges, and where uncertainty about the best acquirer prevents complete contracts. We find that when both technological and commercialization challenges are present, as commonly observed in deep tech innovations, startups are able to capture a smaller fraction of the value created. This introduces a bias in the direction of innovation as projects that are primarily characterized by one type of challenge are more attractive investments compared to projects, equally or more valuable, which face both challenges. We discuss the implications of our model for startup strategies, empirical research and deep tech innovation policies
Extensibility of Association Schemes and GRH-Based Deterministic Polynomial Factoring
No full textThe subject of the present work is the application of the theory of combinatorial schemes to problems in computational algebra. The principal notions of combinatorial schemes which are studied in this work are association schemes (Bannai & Ito (1984), Zieschang (1996, 2005)), m-schemes (Ivanyos, Karpinski & Saxena (2009), Arora et al. (2012)), and presuperschemes (Smith (1994, 2007), Wojdylo (1998, 2001)). The main computational problems considered in this work are polynomial factoring over finite fields, the Schurity problem of association schemes (and its relaxation in the notion of extensibility), and matrix multiplication. We show that each of the latter problems admits a deep connection to the theory of combinatorial schemes, and describe natural algebraic-combinatorial frameworks which capture the essence of their algebraic complexity. As a logical application, we delineate how structural results for combinatorial schemes can translate to fundamental improvements in the realm of computational algebra
Understanding Deep Learning via Analyzing Dynamics of Gradient Descent
No full textThe phenomenal successes of deep learning build upon the mysterious abilities of gradient-based optimization algorithms. Not only can these algorithms often successfully optimize complicated non-convex training objectives, but the solutions found can also generalize remarkably well to unseen test data despite significant over-parameterization of the models. Classical approaches in optimization and generalization theories that treat empirical risk minimization as a black box are insufficient to explain these mysteries in modern deep learning. This dissertation illustrates how we can make progress toward understanding optimization and generalization in deep learning by a more refined approach that opens the black box and analyzes the dynamics taken by the optimizer. In particular, we present several theoretical results that take into account the learning dynamics of the gradient descent algorithm.
In the first part, we provide global convergence guarantees of gradient descent for training deep linear networks under various initialization schemes. Our results characterize the effect of width, depth and initialization on the speed of optimization. In addition, we identify an auto-balancing effect of gradient flow, which we prove to hold generally in homogeneous neural networks (including those with ReLU activation).
In the second part, we study the implicit regularization induced by gradient descent, which is believed to be the key to mathematically understanding generalization in deep learning. We present results in both linear and non-linear neural networks, which characterize how gradient descent implicitly favors simple solutions.
In the third part, we focus on the setting where neural networks are over-parameterized to have sufficiently large width. Through the connection to neural tangent kernels, we perform a fine-grained analysis of optimization and generalization, which explains several empirically observed phenomena. Built on these theoretical principles, we further design a new simple and effective method for training neural networks on noisily labeled data
Interrogation of superior vena cava by deep transgastric transesophageal echocardiography imaging: Clinical applications
No full textThe advantages of intraoperative deep transgastric interrogation by transesophageal echocardiography (TEE) of the superior vena cava (SVC) in comparison to the standard bicaval view was studied in pediatric cardiac surgical cases. The view was found to be helpful in obtaining additional data in pediatric cardiac surgical patients
Domain-agnostic named entity recognition on unstructured text
No full textNamed Entity Recognition (NER) is the task of extracting informing entities belonging to predefined semantic classes from raw text. These semantic classes could be general-purpose like a person, location or domain-specific like genes, protein names in biomedical texts. NER has widespread applications in natural language processing (NLP) and serves as the foundation for applications like question answering, information retrieval and machine translation. Recently, the NER task has got a lot of traction in the research community with the advent of deep learning models like BERT which are able to capture textual semantics very well.
In this work, we present a detailed study approaching the NER task from three different perspectives, namely, sequence labeling, question answering (QA), and span-based classification. We propose a simple span detection and classification pipeline that first detects all mention spans irrespective of entity type and then feeds each mention span as input to a model and expects entity type as output. This setup is the reverse of a traditional QA-based NER system where we feed entity type as input and expect mention spans as output. We also introduce explicit pattern embeddings which compliment character embeddings to learn better word representations even with less training data. Experimental results demonstrate the effectiveness of our proposed domain-agnostic techniques on multiple datasets. We set the new state-of-the-art for BioNLP13CG and give a competitive performance on CoNLL 2003 and JNLPBA datasets. Additionally, we probe into the BERT model and show that mere concatenation of external feature vectors with BERT outputs may not train effectively at the recommended low learning rates for BERT. More sophisticated feature fusion is essential.Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2021-09-16 without embargo termsThe student, Jatin Arora, accepted the attached license on 2021-04-22 at 18:38.The student, Jatin Arora, submitted this Thesis for approval on 2021-04-22 at 19:23.This Thesis was approved for publication on 2021-04-26 at 11:48.DSpace SAF Submission Ingestion Package generated from Vireo submission #16516 on 2021-09-16 at 16:47:19Made available in DSpace on 2021-09-17T01:11:15Z (GMT). No. of bitstreams: 3
ARORA-THESIS-2021.pdf: 378776 bytes, checksum: a1c30756c7e5c8346ec6c657ae66d4a0 (MD5)
JatinArora-MS-CS-Thesis-Source.zip: 154334 bytes, checksum: afb48c4f9b3f14dc8994229fc3c0d588 (MD5)
LICENSE.txt: 4208 bytes, checksum: 494dcebc0812ab8938d846ad76749772 (MD5)
Previous issue date: 2021-04-2
- …
