BEGIN:VCALENDAR VERSION:2.0 PRODID:-//Talks.cam//talks.cam.ac.uk// X-WR-CALNAME:Talks.cam BEGIN:VEVENT SUMMARY:Beyond Flatland: A Geometric Take on Matching Methods for Treatmen t Effect Estimation &\; Efficient World Models with Context-Aware Token ization - Melanie Fernandez Pradier\, Senior Researcher\, MSRC &\; Vinc ent Micheli\, University of Geneva DTSTART:20240911T163000Z DTEND:20240911T180000Z UID:TALK220618@talks.cam.ac.uk CONTACT:Microsoft Research Cambridge Talks Admins DESCRIPTION:THESE TALKS WILL BE PRESENTED AT 198 CAMBRIDGE SCIENCE PARK NO T 21 STATION ROAD\nSpeaker 1:\nAbstract\n: Matching is a popular approach in causal inference to estimate treatment effects by pairing treated and c ontrol units that are most similar in terms of their covariate information . However\, classic matching methods completely ignore the geometry of the data manifold\, which is crucial to define a meaningful distance for matc hing\, and struggle when covariates are noisy and high-dimensional. In thi s work\, we propose\nGeoMatching\n\, a matching method to estimate treatme nt effects that takes into account the intrinsic data geometry induced by existing causal mechanisms among the confounding variables. First\, we lea rn a low-dimensional\, latent Riemannian manifold that accounts for uncert ainty and geometry of the original input data. Second\, we estimate treatm ent effects via matching in the latent space based on the learned latent R iemannian metric. We provide theoretical insights and empirical results in synthetic and real-world scenarios\, demonstrating that GeoMatching yield s more effective treatment effect estimators\, even as we increase input d imensionality\, in the presence of outliers\, or in semi-supervised scenar ios.\n\nSpeaker 2\nAbstract:\nScaling up deep Reinforcement Learning (RL) methods presents a significant challenge. Following developments in genera tive modelling\, model-based RL positions itself as a strong contender. Re cent advances in sequence modelling have led to effective transformer-base d world models\, albeit at the price of heavy computations due to the long sequences of tokens required to accurately simulate environments. In this work\, we propose Δ-IRIS\, a new agent with a world model architecture c omposed of a discrete autoencoder that encodes stochastic deltas between t ime steps and an autoregressive transformer that predicts future deltas by summarizing the current state of the world with continuous tokens. In the Crafter benchmark\, Δ-IRIS sets a new state of the art at multiple frame budgets\, while being an order of magnitude faster to train than previous attention-based approaches.\n\nPlease register:\nhttps://forms.office.com /pages/responsepage.aspx?id=v4j5cvGGr0GRqy180BHbR6ALaDloVjxDrC2QU3AErn5URU RUV1NTQzhKS1VYWktESlA5TUVWR1AzRS4u&route=shorturl LOCATION:Microsoft Research Ltd\, 198 Cambridge Science Park\, Milton Road \, CB4 0AB END:VEVENT END:VCALENDAR