Already Covered

• Basic Neural Network ideas
• Training : Feature creation
• Layers : Different 'levels' of Features

CNN Review

• Pixels in an images are 'organised' :
• • Up/down left/right
  • Translational invariance
• Idea : Use whole image as feature
• • Update parameters of 'Photoshop filters'
• Mathematical term : 'convolution kernel'
• • CNN = Convolutional Neural Network

CNN Filter

Play with a Filter

CNN Model Overview

Image Competition

• ImageNet aka ILSVRC
• over 15 million labeled high-resolution images...
• ... in over 22,000 categories

ImageNet-googlenet

• Pre-trained networks are available :

GoogLeNet (2014)

Layers of Features

AlexNet (2012)

Let's Abuse a CNN

• Last time : Speech Recognition
• Now : Creating Art

Style-Transfer

• Combine :
• • Regular photo (you provide)
  • An artist's style (another photo)
• 'Understand' both using CNN features

Style-Transfer idea

• Start out with a copy of your photo
• Change it to mimic :
• • Fine features from the original
  • Style features from the artist
  • A good looking result

Style-Transfer details

• Start with image==photo (or random noise)
• Update image to minimise (gradient descent) :
• • Pixel-wise L2 vs photo features
  • Gram-matrix similarity vs style features
  • Self total-variational-loss
• Change parameters according to taste

Demo time

github.com/mdda/deep-learning-workshop/
notebooks/2-CNN/6-StyleTransfer/
4-Art-Style-Transfer-inception_tf.ipynb

Some Examples

(Apologies to the artists)

Starry Canal (~ van Gogh)

Mediterranean Canal (~ Picasso )

Cat Cézanne

TF & DL
MeetUp Group

• MeetUp.com / TensorFlow-and-Deep-Learning-Singapore
• Next Meeting :
• • 25-May-2017 @ Google : Topic == "Text"
• Typical Contents :
• • Talk for people starting out
  • Something from the bleeding-edge
  • Lightning Talks

TF Codecamp

• ML Jeju Camp - mentorship from top Google developers
• July 3rd through 28th, 2017 in Jeju, South Korea
• Apply now (limited to 20 people, closes 20-April-2017)
• • Good understanding of TensorFlow and deep learning
  • Submit a project proposal
  • No nationality, gender, age, degree, education requirements
• Release all code written on GitHub

8-week Deep Learning
Developer Course

• July - Sept (catch-up during August)
• Weekly 3-hour sessions will include :
• • Instruction
  • Projects : 3 structured & 2 self-directed
• More information : http://RedCatLabs.com/course
• Expect to work hard...