PyConSG 2016 : Deep Learning Workshop

Presentation Link

I recently lead a 92 minute workshop at PyCon SG 2016 in Singapore.

This workshop was hands-on : After a brief background on deep learning, participants started quickly, interacting with offline experiments with a ConvNet.js model, and also the TensorFlow Playground.

But this on-line portion was partly to allow everyone enough time to get a ~1Gb VirtualBox "appliance" created for the event installed on their laptops. Fortunately, over 90% of the people who came already had VirtualBox installed, which was a huge relief.

Once everyone was up-to-speed tools-wise, the workshop then progressed through a series of Jupyter (fka iPython) notebooks, ranging from Theano basics, through MNIST, to ImageNet networks (pretrained models of both GoogLeNet and Inception-v3 were included in the VM).

Then, for the last half-hour, we went over an interesting application : Reinforcement Learning applied to the game "Bubble Breaker". This application, built from the ground up for PyCon, illustrates how simple it is to get Deep-Q-Learning working - with a 'small board' version being trainable in ~2-3 minutes on participants' laptops. A pretrained full-size model was also included in the VM, which now outperforms its creator...

Naturally, this being a FOSS event, all the source is available on GitHub - if you have questions on the software, please leave an 'issue' there.