Learning-based dimensionality reduction
Abstract
Project Summary: Ever-increasing data sizes impose a direct burden on various stages of the informationpipeline, from the subsequen""t transmission and processing to its final storage and analysis. Unfortunately, therapid rates of data growth have not been complem""ented with matching increases in processing power, and hence,there is an enormous demand for new data acquisition and dimensionalit""y reduction paradigms.In our recent work [1], we introduced a new paradigm, partially addressing the emerging challenge by unifying""compressive sensing with statistical learning theory. Our key idea, which is surprisingly simple in retrospect, is thatby using tr""aining signals and developing training procedures, we can efficiently and effectively learn measurementmatrices that directly acqui""re only the ~relevant~ information during data acquisition. As a result, we can not onlyoutperform the existing state-of-the-art co""mpressive sensing techniques on real-world datasets, but can do so withstrong theoretical guarantees as well as extremely simple an""d efficient linear reconstruction mechanisms. To thisend, we study the following three inter-related research thrusts:Thrust I: No""n-linear decoders and models. This thrust extends our preliminary work, which considers the standardlinear acquisition model along"" with the use of a simple linear decoder, towards non-linear models and reconstructionalgorithms. The mathematical premise of our a"pproach represents a profound re-thinking of the data modelsand dimensionality reduction using continuous optimization algorithms b"ased on convexity, and combinatorialoptimization algorithms based on submodularity.Thrust II: Learning for tasks other than recons""truction. Existing literature has focused almost exclusively onproblems in signal reconstruction, approximation, and estimation in"" noise. However, in many applications, weacquire data for a different final goal, namely, making an inference such as a detection o""r classification decision.As a result, this thrust develops learning-based theory and methods that directly optimize sampling for t"he desiredtasks without necessarily reconstructing the signals involved.Thrust III: Multiple-measurement scenarios. This thrust co"nsiders scenarios in which a signal is simultaneouslymeasured from multiple sensors, which are then combined to perform an inferenc""e task such as estimation.By doing so, we obtain significant performance gains over the single-sensor setting, and present an addit"ionaldegree of flexibility in the overall system design to effectively trade off the available resources.This project will greatly" enhance the applicability of adaptive sampling, overcoming the limitations of conventionaltechniques and simplistic models by deve"loping techniques that learn the required information directlyfrom data. Aside from the broader impact in settings where data compr"ession is relevant, we will pursue a numberof particularly important specific applications, including medical imaging, intercranial"" electro-encephalography,and array signal processing, which relate to the core of C4ISR (Code 31). This project is highly-interdisc""iplinary,and will strengthen connections between the areas of machine learning, signal processing, and optimization, andwill resul"t in new sampling theory and methods as well as new conference and journal publications.Principle Investigator Name: Volkan CevherOrganization: Laboratory for Information and Inference SystemsEcole Polytechnique Federale de LausanneAddress: STI IEL LIONS Stati"on 11, 1015 Lausanne, SwitzerlandEmail: volkan.cevher@epfl.chPhone: +41 216931101Fax: +41 216937600Administrative POC Name: Gosi"a BaltaianOrganization: Laboratory for Information and Inference SystemsEcole Polytechnique Federale de LausanneAddress: STI IEL" LIONS Station 11, 1015 Lausanne, SwitzerlandEmail: gosia.baltaian@epfl.chPhone: +41 216931174Fax: +41 216937600NCAGE Code: SAA8"1DUNS Number: 482271272Total Requested Fundi
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 29, 2017
- Source ID
- N629091712111
Entities
People
- Volkan Cevher
Organizations
- Office of Naval Research
- Swiss Federal Institute of Technology in Lausanne
- United States Navy