Timezone: »

 
Oral
Generative Temporal Models with Spatial Memory for Partially Observed Environments
Marco Fraccaro · Danilo J. Rezende · Yori Zwols · Alexander Pritzel · S. M. Ali Eslami · Fabio Viola

Wed Jul 11 05:30 AM -- 05:50 AM (PDT) @ A7
in Representation Learning 3 »

In model-based reinforcement learning, generative and temporal models of environments can be leveraged to boost agent performance, either by tuning the agent's representations during training or via use as part of an explicit planning mechanism. However, their application in practice has been limited to simplistic environments, due to the difficulty of training such models in larger, potentially partially-observed and 3D environments. In this work we introduce a novel action-conditioned generative model of such challenging environments. The model features a non-parametric spatial memory system in which we store learned, disentangled representations of the environment. Low-dimensional spatial updates are computed using a state-space model that makes use of knowledge on the prior dynamics of the moving agent, and high-dimensional visual observations are modelled with a Variational Auto-Encoder. The result is a scalable architecture capable of performing coherent predictions over hundreds of time steps across a range of partially observed 2D and 3D environments.

[ PDF] 

Author Information

Marco Fraccaro (Technical University of Denmark)
Danilo J. Rezende (DeepMind)
Danilo J. Rezende

Danilo is a Senior Staff Research Scientist at Google DeepMind, where he works on probabilistic machine reasoning and learning algorithms. He has a BA in Physics and MSc in Theoretical Physics from Ecole Polytechnique (Palaiseau – France) and from the Institute of Theoretical Physics (SP – Brazil) and a Ph.D. in Computational Neuroscience at Ecole Polytechnique Federale de Lausanne, EPFL (Lausanne – Switzerland). His research focuses on scalable inference methods, generative models of complex data (such as images and video), applied probability, causal reasoning and unsupervised learning for decision-making.

Yori Zwols (DeepMind)
Alexander Pritzel (Deepmind)
S. M. Ali Eslami (DeepMind)
S. M. Ali Eslami

S. M. Ali Eslami is a staff research scientist at DeepMind working on problems related to artificial intelligence. Prior to that, he was a post-doctoral researcher at Microsoft Research in Cambridge. He did his PhD in the School of Informatics at the University of Edinburgh, during which he was also a visiting researcher in the Visual Geometry Group at the University of Oxford. His research is focused on figuring out how we can get computers to learn with less human supervision.

Fabio Viola (DeepMind)

Related Events (a corresponding poster, oral, or spotlight)

More from the Same Authors