Yes we GAN! The first Machine Learning Tokyo Workshop on GANs was a great success. We held the 1-day-workshop at Tokyo Chapter in Roppongi and 20 machine learning engineers came to hear our three instructors: Mustafa Yagmur works as a machine learning engineer for a Weathernews Inc. in Tokyo. Gregorio Nuevo Castro is currently writing his MSc thesis in Artificial Intelligence and interning at Fujitsu Labs in Tokyo and Dimitris Katsios is a PhD candidate in Artificial Intelligence and interning at Fujitsu Labs in Tokyo as well.
Together we deep-dived into Generative Adversarial Networks. The core principle is generating fake data, that approximates the probability distribution of the training data, let’s call it real data. This procedure requires two networks in a training loop: a Generator and a Discriminator. The Discriminator was trained on the real data. The Generator starts with random noise (z), which is being fed to the Discriminator, a binary classifier, that judges if the data seen is fake (0) or real (1). If the Discriminator returns a probability close to 1, the Generator did a good job generating authentic looking fake data.
Our first simple implementation was demonstrated on a uniform distribution of a Gaussian function. Iterations were neatly visualized in PyTorch which made it really easy to grasp the core concept of how things play together. Then we moved on to a more complex implementation, where we generated handwritten digits, using the MNIST dataset. Our last step was looking at tips and tricks on how to optimize GANs, we made some flags and experimented with the fashion MNIST.
Here you can find the resources we have provided to workshop participants. You will find a bundle of introductory materials, the presentation on the theory of GANs and the technical implementations in PyTorch and Keras.
We look forward to holding next Machine Learning Tokyo workshop. Join the community and stay tuned for updates.
Here are some impressions of the workshop: