From Shallow to Deep Learning for Inverse Imaging Problems - Carola-Bibiane Schönlieb, Cambridge
Alan Turing Institute via YouTube
Overview
Explore the intersection of mathematics and data science in this 36-minute conference talk by Carola-Bibiane Schönlieb from Cambridge University. Delve into inverse imaging problems, progressing from shallow to deep learning techniques. Examine dose-imposed problems, variation regularization, and their applications. Understand key concepts like regularizers, deep neural networks, and bilevel optimization. Learn about pseudoinverse methods, training regularizers, and gradient descent steps in neural networks. Analyze results and takeaway messages, while also considering the disadvantages of these approaches. Gain insights from this cutting-edge research presented at the Isaac Newton Institute's workshop on "Approximation, sampling and compression in data science," which aims to foster collaboration among experts in mathematics, statistics, computer science, and engineering.
Syllabus
Intro
Dose
Imposed problems
Variation regularization
Applications
Examples
Regularizers
The idea
Key words
Deep neural networks
What are people doing
Pseudoinverse
Learning the regularizer
Bilevel optimization problem
Training the regularizer
Gradient descent step
Neural networks
Results
Takeaway messages
Disadvantages
Taught by
Alan Turing Institute