12:00pm - 1:00pm
EB T1 or Zoom: https://binghamton.zoom.us/j/91544656179?pwd=RGpJc1Q0MnZhYmxHQ1BqZlBqcGZpZz09
6:30pm - 8:00pm
Presenter: Senqi Cao
Advisor: Professor Zhongfei (Mark) Zhang
When: Wed, 12/13, 6:30 PM
Zoom Link: https://binghamton.zoom.us/j/97862558439
Title: Advancing Deep Neural Network Reliability: Uncertainty Quantification and Enhanced Training Methods
Abstract: This thesis presents two significant contributions to the field of
Deep Neural Network (DNN) research, addressing both the reliability
and performance of DNN models. The first part of the thesis focuses on
the critical issue of uncertainty quantification in DNNs, a pivotal
aspect in high-stakes applications such as medical diagnostics and
autonomous driving. We introduce a novel uncertainty factorization
formula, which breaks down a typical joint density into three sources
of uncertainty. This theoretical foundation is coupled with a
practical, general and flexible framework for predictive uncertainty
estimation, demonstrating promising results in enhancing DNN
reliability in unfamiliar situations.
The second part of the thesis extends our focus to improving DNN
performance under limited computational and memory resources. It
delves into the unexplored mechanisms behind the effectiveness of
existing techniques such as label smoothing, self-distillation, weight
averaging and deep ensembles. We propose a theoretical framework
offering a unified explanation for the performance gains of these
existing methods. Building on this analysis, we develop a novel DNN
training methodology, composed of three integral techniques:
exploration, sampling, and integration. This method not only
elucidates the principles driving performance improvement in deep
learning models but also outperforms deep ensembles in terms of
prediction accuracy, requiring only a single forward pass, which is
significantly more computationally efficient.
Overall, this thesis contributes to both theoretical understanding and
practical advancements in DNNs, offering new insights and
methodologies that improve the reliability and performance of these
models in real-world applications.
Advisor: Professor Zhongfei (Mark) Zhang
When: Wed, 12/13, 6:30 PM
Zoom Link: https://binghamton.zoom.us/j/97862558439
Title: Advancing Deep Neural Network Reliability: Uncertainty Quantification and Enhanced Training Methods
Abstract: This thesis presents two significant contributions to the field of
Deep Neural Network (DNN) research, addressing both the reliability
and performance of DNN models. The first part of the thesis focuses on
the critical issue of uncertainty quantification in DNNs, a pivotal
aspect in high-stakes applications such as medical diagnostics and
autonomous driving. We introduce a novel uncertainty factorization
formula, which breaks down a typical joint density into three sources
of uncertainty. This theoretical foundation is coupled with a
practical, general and flexible framework for predictive uncertainty
estimation, demonstrating promising results in enhancing DNN
reliability in unfamiliar situations.
The second part of the thesis extends our focus to improving DNN
performance under limited computational and memory resources. It
delves into the unexplored mechanisms behind the effectiveness of
existing techniques such as label smoothing, self-distillation, weight
averaging and deep ensembles. We propose a theoretical framework
offering a unified explanation for the performance gains of these
existing methods. Building on this analysis, we develop a novel DNN
training methodology, composed of three integral techniques:
exploration, sampling, and integration. This method not only
elucidates the principles driving performance improvement in deep
learning models but also outperforms deep ensembles in terms of
prediction accuracy, requiring only a single forward pass, which is
significantly more computationally efficient.
Overall, this thesis contributes to both theoretical understanding and
practical advancements in DNNs, offering new insights and
methodologies that improve the reliability and performance of these
models in real-world applications.
12:00pm - 1:00pm
EB T1 or Zoom. See link below.
