Object Files and Schemata: Factorizing Declarative and Procedural Knowledge in Dynamical Systems

  • 2020-06-29 17:45:03
  • Anirudh Goyal, Alex Lamb, Phanideep Gampa, Philippe Beaudoin, Sergey Levine, Charles Blundell, Yoshua Bengio, Michael Mozer
  • 0

Abstract

Modeling a structured, dynamic environment like a video game requires keepingtrack of the objects and their states (\emph{declarative} knowledge) as well aspredicting how objects behave (\emph{procedural} knowledge). Black-box modelswith a monolithic hidden state often lack \emph{systematicity}: they fail toapply procedural knowledge consistently and uniformly. For example, in a videogame, correct prediction of one enemy's trajectory does not ensure correctprediction of another's. We address this issue via an architecture thatfactorizes declarative and procedural knowledge and that imposes modularitywithin each form of knowledge. The architecture consists of active modulescalled \emph{object files} that maintain the state of a single object andinvoke passive external knowledge sources called \emph{schemata} that prescribestate updates. To use a video game as an illustration, two enemies of the sametype will share schemata but will each have their own object file to encodetheir distinct state (e.g., health, position). We propose to use attention tocontrol the determination of which object files to update, the selection ofschemata, and the propagation of information between object files. Theresulting architecture is a drop-in replacement conforming to the sameinput-output interface as normal recurrent networks (e.g., LSTM, GRU) yetachieves substantially better generalization on environments that havefactorized declarative and procedural knowledge, including a challengingintuitive physics benchmark.

 

Quick Read (beta)

loading the full paper ...