CNNs @ 3IT

Victor Schmidt & Simon Verret

Mila

 

3IT - Université de Sherbrooke

 

14 Novembre 2019

Hier

MLP

 

Loss

 

Backprop

Aujourd'hui

Convolutions

 

Réseaux convolutionnels

 

GANs

Motivation

Classification

Semantic Segmentation

Detection

Instance Segmentation

Image Generation

Image to Image Translation

How to deal with Images?

Filters

Stride = 2

Padding = 1

Convolutional layers

Some more convolutions

Padding = same

Dilated

Transposed

Small translations: pooling

In practice

Convolutional

Neural
Networks

Multi-channel convolutions

Normalizing the inputs

A first network: LeNet

LeCun, 1998

Stabilising  training: Batch Norm

Allows the training of deeper networks

 

Less sensitive to initialization

 

Larger learning rates can be used

 

Faster and more stable convergence

Ioffe, 2015

Updated LeNet

Everything's differentiable!

Some more classification

ImageNet

1000 classes

~10 000 images per classes

1.4M in total

Top-1 ou Top-5 accuracy

AlexNet

First deep networks to win "ImageNet competition"

Krizhevsky, 2012

VGG

Even deeper, more regular kernels (3x3)

Simonyan, 2014

GoogLeNet

"We need to go deeper"

Szegedy, 2014

Depth is not everything

Or is it? : ResNet

He, 2015

SOTA

 Xie, 11 Nov 2019

Transfer Learning

Strategies

When?

Generative Adversarial Networks

Models

Value function

Goodfellow, 2014

Conditioning

Mirza, 2014

Unsupervised Image to Image Translation

Zhu, 2017

Merci!

schmidtv@mila.quebec

htttps://vict0rs.ch

 

More reading material