SIO Faculty Candidate Seminar - Peter Gerstoft


 
10/19/2017 - 4:00pm
SCRIPPS INSTITUTION OF OCEANOGRAPHY OCEAN ACOUSTICS FACULTY CANDIDATE SEMINAR

DATE:          October 19th, Thursday, 4pm  
            (Additionally, a chalk talk will be held on 10/20 at 2pm in NH101) 

LOCATION:     Spiess Hall 330
 
SPEAKER:      Peter Gerstoft, Ph.D.
            SIO
            
TITLE:          Sparse modeling and machine learning in ocean acoustics

 

 
ABSTRACT:    
 
Acoustic and seismic signals are used to infer properties of the Earth and oceans. These signals, which are generated by natural sources as earthquakes and storms (or man-made sources), travel great distances, acquiring environmental information in the process. When these waves are recorded by arrays of sensors, an inverse problem can estimate source location, and ocean/earth structures and properties. These inverse problems are critical for advancing our understanding of the structure of the Earth and oceans, and climate dynamics.

Compressive sensing (CS) have demonstrated superior performance to conventional methods by assuming that high-dimensional signals are simply the linear superposition of few prototype signals from a potentially large set of choices, termed a dictionary. Solutions to such problems are obtained by enforcing that each signal realization consists of few components. Recent work in my NoiseLab research group has with acoustic data demonstrated significant improvements of beamforming and source localization using CS. It has further been shown that dictionaries can be learned directly from the data. This dictionary learning methodology, a form of unsupervised learning, has been applied to ocean acoustic data, and in tomography it has the potential to give improved resolution. We have further developed supervised learning techniques based on neural networks and support vector machines to improve acoustic source localization.

The NoiseLab is focused on developing compressive sensing and machine learning-based techniques to leverage the increasing volume of environmental data, and work across disciplines to improve environmental signal processing methods. Recent data analysis of Arctic and Antarctic noise signals concludes the talk.
 
 
 
Faculty Host:  John Hildebrand (jhildebrand@ucsd.edu)
For more information on this event, contact: 
lcosti@ucsd.edu
Event Calendar: 
Location: 
Spiess 330