Advanced Deep Learning

by Prof. Seungchul Lee
Industrial AI Lab
http://isystems.unist.ac.kr/
POSTECH

Table of Contents

Artistic Style Transfer for Video

  • Vincent van Gogh
In [1]:
%%html
<center><iframe src="https://youtube.com/embed/ckqemfh0JMM?rel=0" 
width="560" height="315" frameborder="0" allowfullscreen></iframe></center>
  • Picasso
In [2]:
%%html
<center><iframe src="https://youtube.com//embed/nMwU4avioVo?rel=0" 
width="560" height="315" frameborder="0" allowfullscreen></iframe></center>

Revisit CNN

  • Hierarchical feature representation

    • Contents representation

    • Style representation


Style Transfer

  • Image construction


2. Discriminant Model vs. Generative Models

Imbalanced Data

  • Not enough data from faulty status

  • Data Imbalance

    • Under sampling
    • Over sampling
    • Re-weighting
    • (Ada)Boosting


Generative Model

  • Data imbalance
    • Problematic in reality
    • For example, 98% OK, 2% NG


Revisit Autoencoder with MNIST Data

Data Generation from Decoder

3. Generative Adversarial Networks (GAN)


Turing test

  • How to generate data?
    • Train through competition
    • Generator vs. Discriminator


4. Deep Learning Implementation

Computation Environment for Model Learning

  • Development environment (open source)
    • Ubuntu 14.04
    • Python3
    • TensorFlow
  • Machine
    • GPU: GeForce GTX TITAN X (PASCAL)
    • CPU: Intel i7-5930k 6 Core 3.5GHz processor


  • Parallel computing
    • Multi GPUs


Implementation of Deep Learning Model


In [3]:
%%html
<center><iframe src="https://www.youtube.com/embed/-eqVRXtX44Y?rel=0" 
width="560" height="315" frameborder="0" allowfullscreen></iframe></center>
In [4]:
%%javascript
$.getScript('https://kmahelona.github.io/ipython_notebook_goodies/ipython_notebook_toc.js')