Join us for the upcoming technical talk —attend in person at Stockman’s Restaurant or virtually via (https://teams.microsoft.com/l/meetup-join/19%3ameeting_MGYxYTkzNmMtMzg0MS00MjdjLTkwYmMtOWNkMWMwNjMzOWFi%40thread.v2/0?context=%7b%22Tid%22%3a%224cf464b7-869a-4236-8da2-a98566485554%22%2c%22Oid%22%3a%2264d770c8-e8c2-4ef1-b2f4-1e7ebd03bd37%22%7d)
Title:
An Introduction to Uncertainty Quantification for Deep Learning-Based Models.
Abstract:
Deep learning has become increasingly popular in recent years, largely due
to its ability to achieve state-of-the-art performance on many complex tasks. However,
even well-trained models cannot detect incorrect predictions (e.g., ChatGPT cannot
detect its own hallucinations), nor do they have the inherent ability to express what they
“do not know”. Uncertainty quantification (UQ) is a recently emerging subfield of deep
learning that provides tools for dealing with such issues. UQ methods estimate the
uncertainty associated with a model's predictions, enabling the detection of
misclassified samples or those from outside the semantic domain of the training data
(i.e., samples far from the training distribution). UQ is crucial for models deployed in
high-risk applications, such as medical diagnostics, materials science, nuclear
engineering, and autonomous vehicles. This presentation gives a basic introduction to
UQ in the context of deep learning and briefly details some important methods that have
been consistently shown to perform well in a variety of tasks.
Speaker: Kyle Lucke
Kyle Lucke is a PhD student in the Computer Science Department at the University of Idaho,
working in the Machine Intelligence and Data Analytics lab. He received his Bachelor of Science
in Computer Science with a minor in Mathematics at the University of Montana in 2019. He
graduated Summa Cum Laude and received the President's Outstanding Senior award. He then
obtained his Master of Science in Computer Science at the University of Montana in 2021,
where he developed novel statistical methods for protein inference. His current research
focuses on developing novel uncertainty quantification approaches for high-stakes deep
learning-based applications. In the past, he has worked on novel approaches for semantic
image segmentation for biomedical images, physics-informed neural networks, and generative
models for cybersecurity applications. His other research interests include robust AI, adversarial
machine learning, and deep learning-based approaches for biomedical image processing.
1175 Pier View Dr, Idaho Falls, Idaho, United States, 83402, Virtual: https://events.vtools.ieee.org/m/490178