[Ml-stat-talks] From Single Cells to Exploding Stars: Machine Learning on All Scales

David Blei blei at CS.Princeton.EDU
Sat Sep 28 16:23:39 EDT 2013


see below for a lunchtime talk next week.



>From Single Cells to Exploding Stars: Machine Learning on All Scales
Dr. Thomas J. Fuchs, California Institute of Technology and Jet
Propulsion Laboratory
EQuad E-219, Thursday 3 October, 12pm (lunchtime seminar)

We are in the midst of a data-driven revolution. The current explosion
of data in nearly all sciences rivals such historic events like the
invention of the printing press or Galileo's first telescope. Unlike
ever before in the history of science these vast amounts of data
exceed the capabilities of even the best domain experts. Thus gaining
new insight into nature is fundamentally a collaborative effort and it
is machine learning which gives us the capabilities to gain knowledge
from big data and it hands us the tools to tackle a multitude of novel
and exciting research questions.
In this talk I will focus on projects in cancer research and astronomy
and I will argue for a joint discriminative bottom-up and generative
top-down approach for parameter estimation and classification. First,
fast and simple features based on low-level visual cues are used to
train discriminative ensemble classifiers. Then, the output of these
models is utilized to restrict the hypothesis space and facilitate
parameter estimation for complex generative models. Adopting a
Bayesian point of view allows us to perform parameter inference and
uncertainty quantification based on the estimated posterior
distributions. In this context a challenge arises from the fact that a
large number of interesting statistical models have no tractable
likelihood. To this end, I will describe an adaptive population Monte
Carlo framework based on Approximate Bayesian Computation (ABC) which
makes likelihood-free inference feasible. The utility and performance
of the proposed approach is demonstrated for applications on
microscopic scale in computational pathology and up to macroscopic
scale in space exploration.

I will conclude the talk with an update on how these recent
breakthroughs in large scale machine learning at JPL facilitate land
coverage classification in a joint project with the Ecohydrology Lab
at Princeton.

Dr. Thomas Fuchs is a research technologist at NASA's Jet Propulsion
Laboratory in Pasadena and visiting scientist at the California
Institute of Technology where he occasionally teaches and organizes
the machine learning seminar. His research focuses on the development
of new ensemble methods and Bayesian sampling techniques for large
scale machine learning. Thomas is PI and CoI for several big data
related research efforts at JPL. Thomas' postdoctoral research at the
computational vision lab of Pietro Perona at Caltech and Larry
Matthies' computer vision group at JPL was focused on approximate
Bayesian computation (ABC) and its application on likelihood-free
parameter estimation for computer vision and medical imaging. In 2010
Thomas received his PhD (Dr.Sc.) from ETH Zurich for his work in the
machine learning laboratory of Joachim Buhmann. During this time he
also completed the Ph.D. program on System Biology and Medicine from
ETH's CC-SPMD. Thomas received his MSc degree (Dipl.-Ing.) from TU
Graz where he majored in technical mathematics with a minor in
computer science. Most of the work for his master thesis on Bayesian
networks was conducted at SCR in Princeton.

More information about the Ml-stat-talks mailing list