One Network for Various Targets

By providing an additional constant controller input which remains time invariant during the generation of some fovea trajectory, various targets can be specified for various trajectories.

The number of $C$'s input units was doubled: For each original input unit there was another input unit whose constant activation defined the desired activation at the end of a fovea trajectory (the goal). (This goal-defining feature is also relevant for `higher-level' sub-goal generating processes to be addressed later.) $M$ remained unchanged, the same parameters as above were used for the training phase.

The controller was able to learn to look for parts of a scene which matched the time invariant input. See figure 5 for an illustration of trajectories leading to different targets in the same scene.

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