BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//132.216.98.100//NONSGML kigkonsult.se iCalcreator 2.20.4//
BEGIN:VEVENT
UID:20260916T123416EDT-4562lvR7Of@132.216.98.100
DTSTAMP:20260916T163416Z
DESCRIPTION:Abstract\n\nAnomaly Detection (AD) is a critical yet challengin
 g task due to the scarcity of abnormal samples. Self-Supervised Learning (
 SSL) offers a promising solution by enabling effective representation lear
 ning from abundant normal data. While SSL-based approaches have achieved s
 ignificant success in image-based AD\, their application to other modaliti
 es\, particularly temporal data\, remains relatively underexplored. This t
 hesis investigates the use of SSL for anomaly detection across increasingl
 y complex data settings\, progressing from basic temporal data to more com
 plex and multidimensional data types\, while addressing the challenges ari
 sing from limited anomalous samples.\n\nWe begin by exploring SSL in one o
 f its most accessible temporal modalities: acoustic signals. By representi
 ng audio as time–frequency images\, we apply contrastive learning with aud
 io-specific augmentations to achieve strong performance in anomalous sound
  detection. This demonstrates that SSL can effectively capture temporal pa
 tterns when the signal is mapped to a suitable feature-based representatio
 n. Building on this insight\, we introduce Deep Autoencoding Support Vecto
 r Data Descriptor (DASVDD)\, a more general\, task-agnostic framework that
  integrates a self-supervised autoencoder with an SVDD constraint. Through
  evaluations across multiple modalities\, we demonstrate the effectiveness
  of this joint optimization strategy while also revealing the limitations 
 of modality-agnostic models when confronted with complex temporal dependen
 cies.\n\nThe need for more specialized modeling of Multivariate Time-Serie
 s leads to the introduction of mVSG-VFP\, a framework designed for sensor-
 based vehicle engine monitoring. By leveraging graph-based modeling\, mVSG
 -VFP captures latent dependencies across multiple interdependent sensors. 
 Notably\, like its predecessors\, this model operates within the represent
 ation space to mitigate the impact of sensor noise. This progression culmi
 nates in ARTA\, an adversarial self-supervised framework designed to opera
 te directly on raw\, high-dimensional time-series signals. Unlike previous
  iterations\, ARTA is inherently insensitive to noise\, bypassing the need
  for intermediate feature extraction while ensuring the detector remains r
 obust and interpretable.\n\nThrough this trajectory\, from simple acoustic
  models to specialized multivariate systems\, this thesis develops a serie
 s of increasingly sophisticated SSL frameworks. These contributions demons
 trate how addressing the specific structural constraints of temporal data 
 expands the applicability of self-supervised learning\, advancing the stat
 e of the art in anomaly detection across multiple challenging real-world d
 omains.\n
DTSTART:20260601T130000Z
DTEND:20260601T150000Z
LOCATION:Room 603\, McConnell Engineering Building\, CA\, QC\, Montreal\, H
 3A 0E9\, 3480 rue University
SUMMARY:PhD defence of Hadi Hojjati – Self-supervised representation learni
 ng for anomaly detection
URL:https://www.mcgill.ca/ece/channels/event/phd-defence-hadi-hojjati-self-
 supervised-representation-learning-anomaly-detection-373111
END:VEVENT
END:VCALENDAR