Galileo: Perceiving Physical Object Properties by Integrating a Physics Engine with Deep LearningGalileo: Perceiving Physical Object Properties by Integrating a Physics Engine with Deep LearningWu, Jiajun and Yildirim, Ilker and Lim, Joseph J. and Freeman, Bill and Tenenbaum, Joshua B.2015
Paper summarynipsreviewsThe authors introduce a novel approach for inferring hidden physical properties of objects (mass and friction), which also allows the system to make subsequent predictions that depend on these properties. They use a black-box generative model (a physics simulator), to perform sampling-based inference, and leverage a tracking algorithm to transform the data into more suitable latent variables (and reduce its dimensionality) as well as a deep model to improve the sampler. The authors assume priors over the hidden physical properties, and make point estimates of the geometry and velocities of objects using a tracking algorithm, which comprise a full specification of the scene that can be input to a physics engine to generate simulated velocities. These simulated velocities then support inference of the hidden properties within an MCMC sampler: the properties' values are proposed and their consequent simulated velocities are generated, which are then scored against the estimated velocities, similar to ABC. A deep network can be trained as a recognition model, from the inferences of the generative model, and also from the Physics 101 dataset directly. Its predictions of the mass and friction can be used to initialize the MCMC sampler.
The authors introduce a novel approach for inferring hidden physical properties of objects (mass and friction), which also allows the system to make subsequent predictions that depend on these properties. They use a black-box generative model (a physics simulator), to perform sampling-based inference, and leverage a tracking algorithm to transform the data into more suitable latent variables (and reduce its dimensionality) as well as a deep model to improve the sampler. The authors assume priors over the hidden physical properties, and make point estimates of the geometry and velocities of objects using a tracking algorithm, which comprise a full specification of the scene that can be input to a physics engine to generate simulated velocities. These simulated velocities then support inference of the hidden properties within an MCMC sampler: the properties' values are proposed and their consequent simulated velocities are generated, which are then scored against the estimated velocities, similar to ABC. A deep network can be trained as a recognition model, from the inferences of the generative model, and also from the Physics 101 dataset directly. Its predictions of the mass and friction can be used to initialize the MCMC sampler.