Methods of Temporal Invariances.

To smoothen the error surface of an attentive vision system as described above, one can impose temporal smoothness constraints on the input units. This can be done by constructing a new error function by adding differences in successive fovea inputs to the final input error observed at the end of a fovea trajectory. (The approach is reminiscent of Jordan's work [2], however, Jordan imposes temporal constraints on the output units.)

The effect is that the system develops a preference for temporal invariances in input space. For attentive vision, such temporal invariances can be caused e.g. by fovea movements that follow edges. Thus an unsupervised element (a search for regularities) is introduced into the learning process. (Trivial temporal invariances obtained by stopping the fovea are excluded by the goal directed part of the complete error function.)

An empirical motivation for introducing an explicit preference for temporal invariances is given by the experimentally observed fact that even without such a predefined preference the system liked to generate fovea trajectories following edges.

Back to Reinforcement Learning page
Back to main page on Learning attentive vision