BEGIN:VCALENDAR VERSION:2.0 PRODID:-//Talks.cam//talks.cam.ac.uk// X-WR-CALNAME:Talks.cam BEGIN:VEVENT SUMMARY:Pizza &\; AI January 2019 - Microsoft Research/University of Ca mbridge DTSTART:20190125T173000Z DTEND:20190125T190000Z UID:TALK118372@talks.cam.ac.uk CONTACT:Microsoft Research Cambridge Talks Admins DESCRIPTION:*Speaker 1* - David Janz\n*Title* - Successor Uncertainties: E xploration and Uncertainty in Temporal Difference Learning \n*Abstract* - Probabilistic Q-learning is a promising approach balancing exploration and exploitation in reinforcement learning. \nHowever\, existing implementati ons have significant limitations: they either fail to incorporate uncertai nty about long-term consequences of actions or ignore fundamental dependen cies in state-action values implied by the~Bellman equation. These problem s result in sub-optimal exploration. As a solution\, we develop Successor Uncertainties (SU)\, a probabilistic Q-learning method free of the aforeme ntioned problems. SU outperforms existing baselines on tabular problems an d on the Atari benchmark benchmark suite. Overall\, SU is an improved and scalable probabilistic Q-learning method with better properties than its p redecessors at no extra cost. \n\n*Speaker 2* - Jan Stuehmer\n*Title* - In dependent Subspace Analysis for Unsupervised Learning of Disentangled Repr esentations\n*Abstract* - Recently there has been an increased interest in unsupervised learning of disentangled representations using the Variation al Autoencoder (VAE) framework. Most of the existing work has focused larg ely on modifying the variational cost function to achieve this goal. These modifications usually include a variable regularization strength paramete r\, which can be hard or impossible to choose in an unsupervised manner.\n We first show that methods like beta-VAE simplify the tendency of variatio nal inference to underfit causing pathological over-pruning and over-ortho gonalization of learned components. Second we propose a complementary appr oach: to modify the probabilistic model with a structured latent prior. Th is prior allows to discover latent variable representations that are struc tured into independent vector spaces. The proposed prior has three major a dvantages: First\, in contrast to the standard VAE normal prior the propos ed prior is not rotationally invariant. This resolves the problem of unide ntifiability of the standard VAE normal prior. Second\, extensive quantita tive and qualitative experiments demonstrate that the prior encourages a d isentangled latent representation which mitigates the need of carefully tu ning the regularization strength parameter and therefore facilitates unsup ervised learning of disentangled representations. Third\, the experiments demonstrate that the prior significantly mitigates the trade-off introduce d by modified cost functions like beta-VAE and TCVAE between reconstructio n loss and disentanglement\, which allows to improve these approaches with respect to both disentanglement and reconstruction quality significantly over the state of the art. LOCATION:Auditorium\, Microsoft Research Ltd\, 21 Station Road\, Cambridge \, CB1 2FB END:VEVENT END:VCALENDAR