
Speaker:

Veniamin Morgenshtern 

Helm.AI, USA 
Date:

Thursday, November 9, 2017 
Place:

USI Lugano Campus, room SI004, Informatics building (Via G. Buffi 13) 
Time:

10:3011:30 


Abstract:

The resolution of all microscopes is limited by diffraction. The observed data is a convolution of the emitted signal with a lowpass kernel, the pointspread function (PSF) of the microscope. The frequency cutoff of the PSF is inversely proportional to the wavelength of light. Hence, the features of the object that are smaller than the wavelength of light are difficult to observe. In singlemolecule microscopy the emitted signal is a collection of point sources, produced by blinking molecules. The goal is to recover the location of these sources with precision that is much higher than the wavelength of light. This leads to the problem of superresolution of positive sources in the presence of noise. I will show that the problem can be solved by using convex optimization in a stable fashion. The stability of reconstruction depends on Rayleighregularity of the signal support, i.e., on how many point sources can occur within an interval of one wavelength. The stability estimate is complemented by a converse result: the performance of the convex algorithm is nearly optimal. I will describe an application, developed in collaboration with the group of Prof. W.E. Moerner, where we use these theoretical ideas to improve data processing in modern microscopes.
Lastly, I will talk about the work I currently do as a senior researcher at Helm.AI, a young Silicon Valley startup. Using sophisticated mathematical modeling, we developed a stateoftheart road lane detector for selfdriving cars. The key innovation is that this system requires no humanlabeled data for training, a very scalable semisupervised learning approach. I will not go into proprietary details of the algorithms.


Biography:

Veniamin is a senior researcher at Helm.AI, working on robust perception algorithms for selfdriving cars. He specializes in extracting information from unlabeled video stream in order to build a reliable road lane detector. His current research is multidisciplinary, using tools from computer vision, robust statistics, signal processing, and deep learning.
Before Helm.AI, Veniamin was a Postdoctoral scholar in the Statistics Department at Stanford University, where he developed and analyzed convex optimization algorithms for signal recovery problems in microscopy and in radar.
Veniamin studied mathematics and computer science at SaintPetersburg State University, Russia, graduating with the Dipl. Math. degree with honors in 2004. He then joined the Communication Technology Laboratory at ETH Zurich, Switzerland, as a research assistant, where he worked on problems of information transmission in radio communication networks. Veniamin graduated from ETH Zurich in 2010, receiving the Dr. Sc. degree. His Ph.D. thesis was awarded with an ETH Medal. From 2010 to 2012, he was a postdoctoral researcher at ETH Zurich.
In 2012, Veniamin received a twoyear Fellowship for Advanced Researchers by the Swiss National Science Foundation. In 2015, the team that Veniamin led received a Second Prize in Thomson Reuters Eikon Text Tagging Challenge, a machine learning and natural language processing competition.


Host:

Prof. Antonio Carzaniga 


Faculty of Informatics
Università della Svizzera italiana
Via Giuseppe Buffi 13
CH6904 Lugano
Tel.: +41 (0)58 666 46 90
Fax: +41 (0)58 666 45 36
Email: decanato.inf@usi.ch
Web: www.inf.usi.ch
Twitter: @USI_INF



