Deep Decoder: Concise Image Representations from Untrained Networks - Lecture 2

Overview

Explore a lecture on deep decoders and concise image representations from untrained networks. Delve into models for natural images, including wavelets, sparsity, sparse coding, and neural networks trained on large datasets. Examine how untrained neural nets can serve as models for natural images, focusing on the deep decoder architecture. Learn about image compression techniques and how the deep decoder compares to more complex neural network architectures. Investigate the application of deep decoders in solving inverse problems, particularly in image denoising. Understand the theoretical foundations behind why deep decoders work effectively, including their ability to fit limited noise. Compare the deep decoder's performance to other state-of-the-art methods like the deep image prior. Conclude with insights on how linear upsampling, ReLUs, and linear combinations can efficiently synthesize images, followed by a summary and Q&A session.

Syllabus

Deep Decoder: Concise Image Representations from Untrained Networks Lecture 2
Recovering images from few data requires a model for natural images
Models for Natural Images: Wavelets + Sparsity
Models for Natural Images: Sparse Coding
Models for natural images: neural nets trained on large datasets
This talk: Untrained neural nets as a model of natural images
The deep decoder
Compression
Image compression
In contrast to deep decoder, other neural net architectures are complicated
Solving inverse problems with the deep decoder
Inverse problem
Image recovery with models
Denoising performance
Deep decoder is on par with state of the art for denoising
Why does the deep decoder work?
Why does the deep decoder denoise so well?
The deep decoder
Theory: Deep Decoder can only fit so much noise
Denoising rates
Proof
Deep image prior [ Ulyanov et al., '18]
Comparison to denoising with deep image prior [Ulyanov et al., '18]
How can linear upsampling, ReLUs, and liner combinations synthesize images efficiently?
Summary
Q&A