Latent Prediction Loss
A prediction error computed in the latent space output by a model encoder rather than in pixel space, typically the mean squared error or negative log-likelihood between latent states, which avoids modeling irrelevant factors such as texture and lighting.
Timeline (last 90 days)
PhysiCalWorld is trained end-to-end using only a latent prediction loss, without any manually annotated physical parameters.
Potential predictive loss refers to prediction error computed in latent space rather than pixel space, avoiding the modeling burden of physically irrelevant factors such as texture and lighting.
All Facts (2)
PhysiCalWorld is trained end-to-end using only a latent prediction loss, without any manually annotated physical parameters.
50%UnverifiedPotential predictive loss refers to prediction error computed in latent space rather than pixel space, avoiding the modeling burden of physically irrelevant factors such as texture and lighting.
50%