Fwd: Thesis Defense - May 3, 2023 - Kin Gutierrez Olivares - Applied Mathematics of the Future

[Artur Dubrawski]

Team, please mark your calendars for this intellectual feast brought to us
generously by Kin.

Cheers,
Artur

PS The first week of May will be loaded with events for us. Expect at least
one more phd thesis defense plus at least two new thesis proposals coming
out of our Lab that week!


---------- Forwarded message ---------
From: Diane Stidle <stidle at andrew.cmu.edu>
Date: Mon, Apr 17, 2023, 3:41 PM
Subject: Thesis Defense - May 3, 2023 - Kin Gutierrez Olivares - Applied
Mathematics of the Future
To: ml-seminar at cs.cmu.edu <ML-SEMINAR at cs.cmu.edu>, <stine at wharton.upenn.edu>,
<robertastine at icloud.com>


*Thesis Defense*

Date: May 3, 2023
Time: 12:30pm (EST) - Remote only presentation
PhD Candidate: Kin Gutierrez Olivares

*Title: **Applied Mathematics of the Future*

Abstract:
Novel learning algorithms have enhanced our ability to acquire knowledge
solely from past observations of single events to learn from the
observations of several related events. This ability to leverage shared
useful information across time series is causing a paradigm shift in the
time-series forecasting practice. Still, machine learning-based forecasting
faces pressing challenges that limit its usability, usefulness, and
attainable real-world impact, including human interpretability, the ability
to leverage structured information, generalization capabilities, and
computational costs. This thesis tackles these challenges by bridging the
gap between machine learning and classic statistical forecasting methods.

We organized the thesis as follows. We introduce the time-series
forecasting task, accompanied by a short review of modern forecasting
models, their optimization, and forecast evaluation methods. In the
following chapters, we present our approach with three case studies. First,
we augment state-of-the-art neural forecasting algorithms with
interpretability capabilities inspired by time series decomposition
analysis; we illustrate its application in the short-term electricity price
forecasting task. Second, we improve neural forecasting generalization and
computational efficiency in the long-horizon setting through a novel
wavelet-inspired algorithm that assembles its predictions sequentially,
emphasizing components with different frequencies and scales. Third, we
tackle the hierarchical forecasting task, a regression problem with linear
aggregation constraints, by augmenting neural forecasting architectures
with a specialized probability mixture capable of incorporating the
aggregation constraints in its construction. Our approach improves upon the
current state-of-the-art in each of the considered domains.

*Thesis Committee:*
Artur Dubrawski (Chair)
Barnabas Poczos
Russ Salakhutdinov
Robert A. Stine (University of Pennsylvania)

Link to the draft document:
https://drive.google.com/drive/folders/1PZX-1cFVy83M5McOL16WPHIUJYeHEcxn?usp=share_link

Zoom meeting link:
https://cmu.zoom.us/j/96712150727?pwd=UXJIOGorbXZjbEtKdWVPdEFTV3ZPUT09

-- 
Diane Stidle
PhD Program Manager
Machine Learning Department
Carnegie Mellon Universitystidle at andrew.cmu.edu
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/autonlab-users/attachments/20230417/d707bbca/attachment.html>