BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//132.216.98.100//NONSGML kigkonsult.se iCalcreator 2.20.4//
BEGIN:VEVENT
UID:20260524T000324EDT-99127u9bv0@132.216.98.100
DTSTAMP:20260524T040324Z
DESCRIPTION:Abstract\n\nThis thesis explores how knowledge transfer techniq
 ues can enhance both learning and unlearning in deep neural networks. The 
 work begins by investigating knowledge distillation (KD)\, with a focus on
  the under-explored role of the capacity gap between teacher and student m
 odels. Through empirical studies\, we show that tailoring teacher-student 
 capacities to data difficulty improves performance. Inspired by these find
 ings\, a novel KD method is developed that dynamically adjusts teacher cap
 acity\, outperforming existing approaches. The focus then shifts to machin
 e unlearning\, a process enabling models to forget specific data points\, 
 which is critical for applications such as data privacy\, copyright concer
 ns\, and the removal of harmful information. To address the limitations of
  existing exact and approximate unlearning methods\, this thesis introduce
 s transfer unlearning\, which replaces data flagged for future removal wit
 h carefully selected auxiliary samples. This approach avoids retraining\, 
 ensures model performance\, and provides practical unlearning guarantees. 
 Finally\, this thesis introduces the hypothesis that residual information 
 left in internal layers can undermine the effectiveness and security of un
 learning methods. To address this\, a method inspired by domain adaptation
  is proposed\, shifting unlearning to the representation space using domai
 n adversarial training. This method aligns the unlearned model’s internal 
 representations with those of an oracle model trained without the data to 
 be forgotten. Empirical results demonstrate that the proposed approaches o
 utperform competing methods across benchmarks and strengthen defenses agai
 nst both black-box and white-box attacks\, advancing the state of unlearni
 ng in deep neural networks.\n
DTSTART:20250528T180000Z
DTEND:20250528T200000Z
LOCATION:Room 603\, McConnell Engineering Building\, CA\, QC\, Montreal\, H
 3A 0E9\, 3480 rue University
SUMMARY:PhD defence of Naz Mohammadi Sepahvand – Knowledge Transfer for Lea
 rning and Unlearning in Deep Neural Networks
URL:/ece/channels/event/phd-defence-naz-mohammadi-sepa
 hvand-knowledge-transfer-learning-and-unlearning-deep-neural-networks-3654
 48
END:VEVENT
END:VCALENDAR