[Pycon] [new paper] "meng shang" - Forge signatures with a GAN

[info a pycon.it]

Title: Forge signatures with a GAN
Duration: 45 (includes Q&A)
Q&A Session: 15
Language: en
Type: Talk

Abstract: - the anatomy of a neural net 
- how does deep learning learn 
- where do we get data to learn from? 
- lets build a simple GAN (generative adversarial net) the same network behind alpha go zero using jupyter notebook 
- run the GAN to generate human like hand written digits based on the MNIST open data set. - where to go (deeper) from here (how to build a open-source/free ML curriculum for yourself) 
- Q&A Here are 2 gif that shows the results: https://cl.ly/0m1C212r1X3w/Screen%20Recording%202018-01-01%20at%2012.13%20PM.gif https://cl.ly/1I3J1h020s09/Screen%20Recording%202018-01-01%20at%2012.15%20PM.gif 
tl:dr of the gifs:
- trained the discriminator net and generator net together for 250 iterations on MNIST dataset. 
- printed the generator’s result ever 5 iterations to show “learning progress” - in the beginning it’s just random scatter plots 
- in approx. 100 iterations we can see the network generate simple digits like 1s and 0s 
- near the end of 250 iterations, we can start to see the generator produce realistic more complex digits like 4, 6, 9s effectively 
- in approx 500 iterations (not shown in the gif) the results are very realistic to the point where it could fool a human discriminator

Tags: [u'neural network', u'Python', u'tensorflow', u'Jupyter', u'machine-learning', u'Deep-Learning', u'datascience', u'Artificial Intelligence', u'pydata']