deep-learning-figures
deep-learning-figures copied to clipboard
Figures I made during my PhD in Deep Learning, for my models and for context
Deep Learning Figures
These figures have been made mostly during my PhD. Some present general concepts / models of Deep Learning, most are to describe the papers I worked on. The source is a PPTX file containing all the figures. You can try to adapt them to your needs if you feel up for it. If so, I recommend to install the fonts that I used.
These figures are used in my PhD thesis if you want to see them used in context and want a full legend.
How I use PowerPoint figures in my LaTeX documents?
For each paper, I have an images/figures.pptx file that contains all my PowerPoint figures. Regularly, I export this PowerPoint into a images/figures.pdf file. I also have a script images/process_figures.sh and run it after each PDF export:
# Split the PDF into pages
pdfsplit.py figures.pdf
# pdfsplit.py is included in this repo. It is designed for Mac. For Linux or Windows, you can find equivalents.
# Remove pages that I keep in the PPTX but I don't actually want to use
rm figures-3.pdf
rm figures-4.pdf
# Compress some pages if needed, when they contain big images, you need
compress_pdf () {
gs -sDEVICE=pdfwrite -dNOPAUSE -dQUIET -dBATCH -dPDFSETTINGS=/printer -dCompatibilityLevel=1.4 -sOutputFile=$1-comp.pdf $1.pdf
mv $1-comp.pdf $1.pdf
}
compress_pdf figures-1 &
compress_pdf figures-2 &
wait
# Remove the write part of each figure's page
for f in `ls figures-*.pdf`; do
pdfcrop $f $f & # pdfcrop came with my latex install. It's this: https://ctan.org/pkg/pdfcrop
done
wait
# Rename into more usable names
mv figures-1.pdf intro_CV.pdf
mv figures-2.pdf intro_ML.pdf
# ...
General figures
Intro of Computer Vision

Intro of Machine Learning

Intro of Neural Nets

Intro of ConvNets

Intro of Disentangling

Famous ConvNets architectures

VGG architecture by [T. Durand](https://github.co

Illustration of Auto-Encoders

Illustration of Denoising Auto-Encoders

Illustrations of Variational Auto-Encoders


Illustration of GAN

Illustration of Ladder Networks

SHADE: Information-Based Regularization for Deep Learning
Goal of the model

Minimizing Entropy

Minimizing Conditional Entropy

HybridNet: Classification and Reconstruction Cooperation for Semi-Supervised Learning
Model overview

Intuition

General architecture

Losses

ConvLarge architecture

Example of architecture

Branch balancing effect

Merge strategies


HybridNet with SHADE

DualDis: Dual-Branch Disentangling with Adversarial Learning
Overview

Architecture

Comparison with other models
