Slide 1

Slide 1 text

Ensemble of GANs as a Data Augmentation Technique for Alzheimer research by Raul Pino PyCon Italia 2023

Slide 2

Slide 2 text

Designing and Developing your own Capstone Project by Raul Pino PyCon Italia 2023

Slide 3

Slide 3 text

Using PyTorch to Generate Demented Brains by Raul Pino PyCon Italia 2023

Slide 4

Slide 4 text

● Alzheimer, motivation, and research. ● Capstone project: Idea + value added + metrics. ● Bases: ○ Intro to Deep Learning. ○ Original DCGAN paper. ○ Ensembles of GANs paper. ● Solution & Demo: ○ Dataset exploration. ○ GAN base architecture. ○ Ensemble techniques implementation. ○ Results. ● Takeaways & beyond. Agenda

Slide 5

Slide 5 text

About me ● Born in Venezuela. ● +10 years of exp as Software Engineer & AI enthusiast (ML Eng recently). ● Living in Chile. ○ Elementus - uBiome - Groupon. ● <3 AI, Coffee, Scuba Diving, …

Slide 6

Slide 6 text

Alzheimer, motivation & research ● Data scarcity. ● Anonymization. ● Sampling Bias. ● Art & awareness. ● Most common case of dementia. ● Neurodegenerative, loss of neurons. ● Difficult diagnosis. ● No cure.

Slide 7

Slide 7 text

Capstone project: idea + value added + metrics Deep Learning & Machine Learning Engineer programs. ● Something related to Alzheimer? Datasets? ● Background research? Previous work? ● Anything new? Any papers? Any interesting mix? ○ Ensembles of GANs? #wat? ● Any metrics to test solution?

Slide 8

Slide 8 text

Basics: Intro Deep Learning Neural Network on structured data like a csv.

Slide 9

Slide 9 text

Basics: Intro Deep Learning Neural Network on structured data like a csv.

Slide 10

Slide 10 text

Basics: Intro Deep Learning Neural Network on structured data like a csv.

Slide 11

Slide 11 text

Basics: Intro Deep Learning Unstructured data like an img. (But what is a neural network? | Chapter 1, Deep learning - 3Blue1Brown)

Slide 12

Slide 12 text

Basics: Intro Deep Learning Supervised Learning

Slide 13

Slide 13 text

Basics: GANs & DCGAN Paper Unsupervised Learning

Slide 14

Slide 14 text

Basics: GANs & DCGAN Paper Architecture guidelines for stable Deep Convolutional GANs: ● Replace any pooling layers with strided convolutions (discriminator) and fractional-strided convolutions (generator). ● Use batchnorm in both the generator and the discriminator. ● Remove fully connected hidden layers for deeper architectures. ● Use ReLU activation in generator for all layers except for the output, which uses Tanh. ● Use LeakyReLU activation in the discriminator for all layers. .

Slide 15

Slide 15 text

Basics: Ensembles of GANs Ensemble learners - Udacity

Slide 16

Slide 16 text

Suggest three different training techniques: ● Standard Ensemble of GANs (eGANs). ● Self-ensemble of GANs (seGANs). ● Cascade of GANs (cGANs). Basics: Ensembles of GANs

Slide 17

Slide 17 text

Standard Ensemble of GANs (eGANs) Basics: Ensembles of GANs

Slide 18

Slide 18 text

Self-ensemble of GANs (seGANs) Basics: Ensembles of GANs

Slide 19

Slide 19 text

Cascade of GANs (cGANs) Basics: Ensembles of GANs

Slide 20

Slide 20 text

Solution & Demo: Dataset exploration Go to https://github.com/p1nox/gan_ensembles/blob/master/data_exploration_viz.ipynb

Slide 21

Slide 21 text

Solution & Demo: GAN base architecture Go to https://github.com/p1nox/gan_ensembles/blob/master/dcgan_control_model.ipynb

Slide 22

Slide 22 text

Solution & Demo: Ensemble techniques implementation Go to https://github.com/p1nox/gan_ensembles/blob/master/data_exploration_viz.ipynb

Slide 23

Slide 23 text

Solution & Demo: Dataset exploration Go to https://github.com/p1nox/gan_ensembles/blob/master/data_exploration_viz.ipynb

Slide 24

Slide 24 text

Takeaways & beyond Stable Diffusion Midjourney v5 01 Brain Imaging Generation with Latent Diffusion Models 02

Slide 25

Slide 25 text

Resources ● Ensemble of Generative Adversarial Networks as a Data Augmentation Technique for Alzheimer research https://github.com/p1nox/gan_ensembles/blob/master/capstone_project_docs/proposal.pdf ● Hello Future (2022) Generative AI: a new approach to overcome data scarcity. https://hellofuture.orange.com/en/generative-ai-a-new-approach-to-overcome-data-scarcity/ ● 3Blue1Brown: “But what is a neural network? | Chapter 1, Deep learning” - https://www.youtube.com/watch?v=aircAruvnKk ● Ian Good fellow at Lex Fridman Podcast (2019) Ian Goodfellow: Generative Adversarial Networks (GANs) | Lex Fridman Podcast #19. https://www.youtube.com/watch?v=Z6rxFNMGdn0&t=3013s ● Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter (2018) GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. https://arxiv.org/pdf/1706.08500.pdf ● Kaggle (2019) Alzheimer's Dataset ( 4 class of Images) - Images of MRI Segementation. https://www.kaggle.com/datasets/tourist55/alzheimers-dataset-4-class-of-images ● Kaggle (2020) I'm Something of a Painter Myself: Use GANs to create art - will you be the next Monet?. https://www.kaggle.com/competitions/gan-getting-started/overview/evaluation ● Can Kocagil (2021) DCGAN paper implementation using PyTorch to generate faces. https://github.com/cankocagil/DCGAN ● Steven Lang (2021) FID score for PyTorch. https://github.com/mseitzer/pytorch-fid ● Udacity Ensembles learners https://www.youtube.com/watch?v=Un9zObFjBH0

Slide 26

Slide 26 text

“mri happy brain“ Grazie