Action Predictability
The ability to reconstruct actions from consecutive latent states. As a training constraint, it forces the encoder to retain state information that is discriminative with respect to actions, thereby improving the learning of intrinsic dynamics and yielding representations in which control-relevant factors are preserved rather than discarded.
Timeline (last 90 days)
The study theoretically derives that when the encoder merges the distinguishable outcomes of different actions under the same state as well as different appearances, a positive lower bound exists on the squared inverse prediction error.
Action predictability training uses the inverse dynamics prediction error as an auxiliary loss, forcing the encoder to preserve action-discriminative state information.
The idea of action predictability is akin to the information bottleneck principle in contrastive learning: when a model needs to complete a certain prediction task, the encoder is forced to retain the minimal sufficient information necessary to complete the task.
All Facts (3)
The study theoretically derives that when the encoder merges the distinguishable outcomes of different actions under the same state as well as different appearances, a positive lower bound exists on the squared inverse prediction error.
50%UnverifiedAction predictability training uses the inverse dynamics prediction error as an auxiliary loss, forcing the encoder to preserve action-discriminative state information.
50%UnverifiedThe idea of action predictability is akin to the information bottleneck principle in contrastive learning: when a model needs to complete a certain prediction task, the encoder is forced to retain the minimal sufficient information necessary to complete the task.
50%