Fwd: Thesis Defense - Oct. 3, 2018 - Kirthevasan Kandasamy - Tuning Hyper-parameters without Grad-students: Scaling up Bandit Optimisation

[Kirthevasan Kandasamy]

Hi everyone,

I am defending next Wednesday.
You are welcome to drop by.

Thanks!
Samy

---------- Forwarded message ---------
From: Diane Stidle <stidle at andrew.cmu.edu>
Date: Mon, Sep 24, 2018 at 4:19 PM
Subject: Thesis Defense - Oct. 3, 2018 - Kirthevasan Kandasamy - Tuning
Hyper-parameters without Grad-students: Scaling up Bandit Optimisation
To: ml-seminar at cs.cmu.edu <ML-SEMINAR at cs.cmu.edu>, zoubin at eng.cam.ac.uk <
zoubin at eng.cam.ac.uk>, <zoubin at uber.com>


*Thesis Defense*

Date: October 3, 2018
Time: 12:30pm (EDT)
Place: 8102 GHC
PhD Candidate: Kirthevasan Kandasamy

*Title: **Tuning Hyper-parameters without Grad-students: Scaling up Bandit
Optimisation*

Abstract:
This thesis explores scalable methods for adaptive decision making under
uncertainty, where the goal of an agent is to design an experiment, observe
the outcome, and plan subsequent experiments to achieve a desired goal.
Typically, each experiment incurs a large computational or economic cost,
and we need to keep the number of experiments to a minimum. Many of such
problems fall under the bandit framework, where each experiment evaluates a
noisy function and the goal is to find the optimum of this function. A
common use case for the bandit framework, pervasive in many industrial and
scientific applications, is hyper-parameter tuning, where we need to find
the optimal configuration of a black-box system by tuning the several knobs
which affect the performance of the system. Some applications include
statistical model selection, materials design, optimal policy selection in
robotics, and maximum likelihood inference in simulation based scientific
models. More generally, bandits are but one class of problems studied under
the umbrella of adaptive decision-making under uncertainty. Problems such
as active learning and design of experiments are other examples of adaptive
decision-making, but unlike bandits, progress towards a desired goal is not
made known to the agent via a reward signal.
With increasingly expensive function evaluations and demands to optimise
over complex input spaces, bandit optimisation tasks face new challenges
today. At the same time, there are new opportunities that have not been
exploited previously. We study the following questions in this thesis to
enable the application of bandits and more broadly adaptive decision-making
to modern applications.
- Conventional bandit methods work reliably in low dimensional settings,
but scale poorly with input dimensionality. Scaling such methods to high
dimensional inputs requires addressing several computational and
statistical challenges.
- In many applications, an expensive function can be cheaply approximated.
We study techniques that can use information from these cheap lower
fidelity approximations to speed up the overall optimisation process.
- Conventional bandit methods are inherently sequential. We study
parallelisation techniques so as to deploy several function evaluations at
the same time.
- Typical methods assume that a design can be characterised by a Euclidean
vector. We study bandit methods on graph-structured spaces. As a specific
application, we study neural architecture search, which optimises for the
structure of the neural network by viewing as a directed graph with node
labels and node weights.
- Many methods for adaptive decision-making are not competitive with human
experts. Incorporating domain knowledge and human intuition about specific
problems may significantly improve practical performance.
We first study the above topics in the bandit framework and then study how
they can be extended to broader decision-making problems. We develop
methods with theoretical guarantees which simultaneously enjoy good
empirical performance. As part of this thesis, we also develop an open
source platform for scalable and robust bandit optimisation.

Thesis Committee:
Barnabás Póczos(Co-chair)
Jeff Schneider (Co-chair)
Aarti Singh
Zoubin Ghahramani (University of Cambridge)

*Link to draft document: *http://www.cs.cmu.edu/~kkandasa/docs/thesis.pdf

-- 
Diane Stidle
Graduate Programs Manager
Machine Learning Department
Carnegie Mellon Universitystidle at cmu.edu
412-268-1299
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/autonlab-research/attachments/20180924/9ae62bc7/attachment.html>