ArtNet - IBM OpenPOWER Cognitive Cup Contest Winning talk

ArtNet - IBM OpenPOWER Cognitive Cup Contest Winning talk

Transcript

1. ArtNet
   The new age Art Connoisseur ;)
   OpenPower Cognitive Cup Winning Entry
   

   View Slide

2. Developer Challenge
   Experience
   Praveen Sridhar
   @psbots
   

   View Slide

3. About me
   Praveen Sridhar
   @psbots
   Machine Learning R&D for
   Recently joined as
   Machine Learning Engineer at
   

   View Slide

4. Cognitive Cup Challenge
   Problem
   Can computers think “deeply” :P about art?
   

   View Slide

5. What ArtNet Does
   ArtNet “knows” what a painting is about.
   Be it a landscape or still life or a portrait,
   or any of the different art genres.
   Examples of genres :
   Still Life Cityscape Religious
   

   View Slide

6. Implementation
   ArtNet uses
   • a Convolution Neural Network to figure
   out patterns in a painting
   • and classifies it into different genres
   

   View Slide

7. Implementation
   It uses the Keras Deep Learning library to
   train the CNN model
   

   View Slide

8. Implementation
   The OpenPOWER
   Deep Learning Distribution
   Frameworks like Theano and Tensorflow
   available as Pre-built binaries optimized
   for GPU acceleration
   Adding Keras was a breeze, since it runs
   on top of Theano or Tensorflow
   

   View Slide

9. Hyperparameter Optimization
   The Real Deal
   HYPER WHAT?
   

   View Slide

10. Hyperparameter Optimization
    The Real Deal
    Enter the “hyperas” library
    A very simple convenience wrapper around hyperopt for fast
    prototyping with keras models.
    

    View Slide

11. Enter the “elephas” library
    Distributed Deep Learning with Keras & Spark
    Thought : SuperVessel Cloud makes it a breeze to
    spin up multi node spark instances, why not use it as
    Nitrox! ;)
    

    View Slide

12. Going Forward
    • Predict the artist given a painting
    • Help Social Science people in
    understanding art influences :
    ✴ which artist influenced whom?
    ✴ to what extent?
    

    View Slide

13. Thanks!
    Praveen Sridhar
    @psbots
    

    View Slide