Fwd: Thesis Proposal - 3/10/17 - Junier Oliva - Distribution and Histogram (DisH) Learning

[Junier Oliva]

ICYMI I'll be proposing my thesis this Friday! Feel free to come by :)

Best,
Junier

---------- Forwarded message ----------
From: Diane Stidle <diane+ at cs.cmu.edu>
Date: Fri, Mar 3, 2017 at 4:30 PM
Subject: Thesis Proposal - 3/10/17 - Junier Oliva - Distribution and
Histogram (DisH) Learning
To: "ml-seminar at cs.cmu.edu" <ML-SEMINAR at cs.cmu.edu>, Le Song <
lsong at cc.gatech.edu>


*Thesis Proposal*

Date: 3/10/17
Time: 3:00pm
Place: 8102 GHC
Speaker: Junier Oliva

Title: Distribution and Histogram (DisH) Learning

Abstract:
This thesis advances the explicit use of distributions in machine learning.
We develop algorithms that consider distributions as functional
covariates/responses, and methods that use distributions as internal
representations. We consider distributions since they are a straightforward
characterization of many natural phenomena and provide a richer description
than simple point data by detailing information at an aggregate level. Our
approach may be seen as addressing two sides of the same coin: on one side,
we use traditional machine learning algorithms adjusted to directly operate
on inputs and outputs that are probability functions (and sample sets); on
the other side, we develop better estimators for traditional vector data by
making use of, and adjusting internal distributions.

We begin by developing algorithms for traditional machine learning tasks
for the cases when one's input (and/or possibly output) is not a finite
point, but is instead a distribution, or sample set drawn from a
distribution. We develop a scalable nonparametric estimator for regressing
a real valued response given an input that is a distribution, a case which
we coin distribution to real regression (DRR). Furthermore, we extend this
work to the case when both the output response and the input covariate are
distributions; a task we call distribution to distribution regression
(DDR). Moreover, we propose flexible and scalable techniques for
conditional density estimation where one regresses an output response that
is a distribution given a real valued covariate.

After, we look to expand the versatility and efficacy of traditional
machine learning tasks through novel methods that operate with implicit or
latent distributions. Take for example kernel methods that use a
shift-invariant kernel. Here, one's kernel uniquely determines a
distribution, the spectral density, that controls the frequencies
considered over inputs. We show that one may improve the performance of
kernel learning tasks by learning this spectral distribution in a
data-driven fashion using Bayesian nonparametric techniques. Furthermore,
we recast classification as a task on distributions. Namely, the Bayes
classification risk is minimized when the distributions of features of
instances from each particular class are non-overlapping. Hence, we propose
a distribution based task with ties to a Bayes risk to perform supervised
feature learning.

Leveraging the high-level, aggregate information provided by distributions
in these algorithms allows us to improve performance in a broad range of
domains including: cosmology, neuroscience, computer vision, and natural
language processing. Furthermore, the scalable nature of our algorithms are
such that we may scale to millions, even billions of instances.

Thesis Committee:
Barnabas Poczos (Co-Chair)
Jeff Schneider (Co-Chair)
Ruslan Salakhutdinov
Le Song (Georgia Institute of Technology)

Link to draft document:
https://www.dropbox.com/s/tbxj4v0oy35omky/Thesis_Proposal.pdf?dl=0

-- 
Diane Stidle
Graduate Programs Manager
Machine Learning Department
Carnegie Mellon Universitydiane at cs.cmu.edu412-268-1299 <(412)%20268-1299>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/autonlab-users/attachments/20170308/62778b17/attachment.html>