Published on

Deep Learning DIY


I recently lead a 1 hour workshop at FOSSASIA 2018 in Singapore.

Similar to the previous year, this workshop was hands-on : It included a couple of examples in Javascript (rebuilt to run locally, to avoid network problems), followed by the use of a VirtualBox pre-configured linux image (somthing that has proven to be a good way of getting everyone in a workshop up-and-running with the minumum of hassle, and the maximum amount of preinstalled notebooks, models and data - with zero interaction with the WiFi / internet).

Using the updated VirtualBox VM-on-a-stick, we had a quick look at a working NASNet model. And then went on showing how transfer learning can work. That essentially covered the first two sections of the talk : Learning with lots of data; and Learning with some data.

For the second half of this FOSSASIA workshop, I introduced meta-learning, with the emphasis on one-shot learning, to show how models can be learned from very little data. This topic was prompted by the publication (3 weeks earlier) of the OpenAI "Reptile" paper, and the VM included an updated version of the reptile-sines code.

I also (re-)created a stand-alone Reptile one-shot learning demo, so that people could get an intuitive understanding of what a test example for the meta-learning task look like, without being connected to the internet (and without needing the VirtualBox VM running).

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.

Presentation Screenshot

If there are any questions about the presentation please ask below, or contact me using the details given on the slides themselves.

Presentation Content Example

The presentation was kindly recorded by