next up previous
Next: CONCLUSIONS Up: ONGOING AND FUTURE RESEARCH Previous: Methods of Temporal Invariances.

Implications for Learning Selective Attention in the General Case: An Outlook

The system described above (which learns by using the principle of system realization) as well as Whitehead and Ballard's system (mentioned in section 1) can be viewed as implementing selective attention by some sort of external feedback. The system described in [14], which implements `curiosity' and `boredom' by means of adaptive dynamic attention depending on the amount of a model network's ignorance about the external dynamics, also is based on external feedback.

A generalization of the method described above would work as follows. For each input unit of some controller introduce an output unit that gates the current activation of the corresponding input unit at every time step. Train a model network to predict the context dependent effects of suppressing certain input units and emphasizing others. Use system realization as above for learning dynamic selective attention to those input units that are relevant in the context of the current goal.


next up previous
Next: CONCLUSIONS Up: ONGOING AND FUTURE RESEARCH Previous: Methods of Temporal Invariances.
Juergen Schmidhuber 2003-02-21

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