Conditional Mean Embeddings and What Deep Models Don't Know


Details
Summer or not, for this month's meetup we welcome two speakers from UCL and Cambridge to talk about their research. We follow the usual agenda:
-
18:30: doors open, pizza, beer, networking
-
19:00: First talk
-
20:00: Break & networking
-
20:15: Second talk
-
21:30: Close
RSVPs open July 30 4:00pm. As always these events are very popular so please update your RSVP if you can no longer make it.
--London ML Team
• Conditional Mean Embeddings for Reinforcement Learning - John Shawe-Taylor
Conditional Mean Embeddings (CME) provide a way of learning to estimate expectations under unknown distributions. We consider their application to learning the system dynamics for Markov Decision Processes (MDPs). This results in a model-based approach to their solution that reduces the planning problem to a finite (pseudo-) MDP exactly solvable by dynamic programming. Unfortunately the size of the finite MDP scales badly with the amount of experience. By approximating the loss function of the CME the size of the induced (pseudo-) MDP can be compressed while maintaining performance guarantees. At the same time the CME model can itself be approximated using a fast sparse-greedy kernel regression. The performance of the composite method compares favourably with the state-of-the-art methods both in accuracy and efficiency.
Bio: John Shawe-Taylor has contributed to a number of fields ranging from graph theory through cryptography to statistical learning theory and its applications. However, his main contributions have been in the development of the analysis and subsequent algorithmic definition of principled machine learning algorithms founded in statistical learning theory. More recently he has worked on interactive learning and reinforcement learning. He has published over 300 papers with over 42000 citations. Two books co-authored with Nello Cristianini have become standard monographs for the study of kernel methods and support vector machines and together have attracted 21000 citations. He is Head of the Computer Science Department at UCL where he has overseen a significant expansion and witnessed its emergence as the highest ranked Computer Science Department in the UK in the recent 2014 UK Research Evaluation Framework (REF).
• What My Deep Model Doesn't Know... - Yarin Gal
If you give me pictures of various dog breeds – and then you ask me to classify a new dog photo – I should return a prediction with rather high confidence. But if you give me a photo of a cat and force my hand to decide what dog breed it belongs to – I better return a prediction with very low confidence. Similar problems occur in medicine as well: you would not want to falsely diagnose a patient just because your model is uncertain about its output. Even though deep learning tools have achieved tremendous success in applied machine learning, these tools do not capture this kind of information. In this talk I will explore a new theoretical framework casting dropout training in neural networks as approximate Bayesian inference. A direct result of this theory gives us tools to model uncertainty with dropout networks – extracting information from existing models that has been thrown away so far. The practical impact of the framework has already been demonstrated in applications as diverse as deep reinforcement learning, camera localisation, and predicting DNA methylation, applications which I will survey towards the end of the talk.
Bio: Yarin is a Research Fellow in Computer Science at St Catharine's College at the University of Cambridge, and part-time Fellow at the Alan Turing Institute, the UK's national institute for data science. He will obtain his PhD from the Cambridge machine learning group, working with Prof Zoubin Ghahramani and funded by the Google Europe Doctoral Fellowship. Prior to that he studied at Oxford Computer Science department for a Master's degree under the supervision of Prof Phil Blunsom.

Sponsors
Conditional Mean Embeddings and What Deep Models Don't Know