August 27, 2014 · 7:00 PM
This location is shown only to members
We'll check out recent ICML's papers. I picked works that dialog with things we've discussed. Feel free to add suggestions.
We need volunteers to help us lead a session - please send me a message if you care to assist.
1) A gentle dive into deep-learning - we already touched the surface of deep-learning. This is a chance to understand a solution that achieves the state-of-the-art in scene labelling tasks. (link: http://bit.ly/1ltupiC)
2) Can we "recycle" pre-trained (deep learnt) networks? I think this is an interesting question, here is a relevant blog-post (http://bit.ly/1ltuQJN) and here is a corresponding recent ICML paper (http://bit.ly/1v3ovOK). **there are a few more relevant papers - ping if you are interested..
3) The half-way point between supervised learning and reinforcement learning.
Let's discuss the interesting Contextual Bandits problem setting. (link: http://bit.ly/1o7Niwr)
4) Engineering a click-through-rate predictor. The following paper will give us a chance to discuss an online advertising core task and several standard components that compose a solution. (link: http://bit.ly/1APq1Vf)