Fwd: Reminder - Thesis Defense - April 2, 2024 - Cristian Challu - Representing Time: Towards Pragmatic Multivariate Time Series Modeling

[Artur Dubrawski]

Team,

I saw Cristian practice his talk twice already, but I did not have enough
yet!
I am really looking forward to his final presentation tomorrow, it will be
fun and worth your time.

See you all there tomorrow.

Cheers
Artur

---------- Forwarded message ---------
From: Diane Stidle <stidle at andrew.cmu.edu>
Date: Mon, Apr 1, 2024 at 9:44 AM
Subject: Reminder - Thesis Defense - April 2, 2024 - Cristian Challu -
Representing Time: Towards Pragmatic Multivariate Time Series Modeling
To: ml-seminar at cs.cmu.edu <ML-SEMINAR at cs.cmu.edu>, <ywu at stat.ucla.edu>


*Thesis Defense*

Date: April 2, 2024
Time: 11:00am (EDT)
Place: NSH 3305 & Remote
PhD Candidate: Cristian Challu

*Title: Representing Time: Towards Pragmatic Multivariate Time Series
Modeling*

Abstract:
Time series models are specialized in learning temporal dependencies among
observations and interactions between features in a data stream. During the
last decade, the unprecedented success of deep learning models on Computer
Vision and Natural Language Processing has steadily permeated to time
series tasks. From Recurrent Neural Networks to Transformers, new
advancements in architectural design improved capabilities and performance.
Despite this success, I identify several challenges to adopting current
state-of-the-art methods, including handling distribution shifts and
missing data, computational complexity, and interpretability.
The success of DL models is usually explained by their ability to
automatically discover helpful data representations. Multivariate time
series models involve high-dimensional objects with numerous time series
and temporal observations. However, they often exhibit strong temporal
dependencies and inter-feature relations. In this thesis, I propose to
design algorithms for forecasting and anomaly detection tasks that leverage
these dependencies to induce efficient learning of representations that
satisfy desirable properties that can (i) improve the models' performance,
(ii) improve robustness by favoring domain adaptation, and (iii) reduce
overparameterization to improve scalability. The completed work includes
three parts, presenting seven models and algorithms that achieve SoTA
performance in various tasks while addressing key adoption challenges.

In the first part, I explore the dynamic latent space principle and design
latent temporal representations to make robust anomaly detection and
forecasting models. In the second part, I present a novel scalable and
interpretable model for multi-step forecasting based on a non-linear
frequency decomposition with connections to Wavelet theory. It also
features two extensions on using multivariate exogenous covariates for
high-impact domains, including energy and healthcare. Finally, in the third
part, I present a large-scale study on enabling conditions, on both model
design and data characteristics, for transferability of pre-trained models
on time series tasks.

*Thesis Committee:*
Artur Dubrawski, Chair
Roni Rosenfeld
Barnabas Poczos
Ying Nian Wu (UCLA)

Link to Draft Document:
https://drive.google.com/file/d/17luIWYw3bLRYe-BuXn1QstlQad0ozXu8/view?usp=sharing

Zoom meeting link:
https://cmu.zoom.us/j/99919622332?pwd=Y0E0ZU0xMCswbmFyTXBzNDRuK05mZz09

Meeting ID: 999 1962 2332
Passcode: 902254

-- 
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/20240401/4a00dce7/attachment.html>