next up previous
Next: Bibliography Up: LEARNING TO GENERATE ARTIFICIAL Previous: Implications for Learning Selective


Previous approaches to pattern recognition with neural networks emphasized the parallel `static' aspects of information processing. However, even in apparently static domains as target detection in stationary environments much can be gained by introducing sequential elements and dynamic selective attention.

This paper demonstrates that the principle of system realization and gradient descent through a model network can be used for learning certain cases of dynamic selective attention. The context is given by attentive vision: It is demonstrated that an imperfect model network which emulates the fovea dynamics can contribute for learning perfect solutions to certain target detection problems. However, the concept of system realization is general enough to allow less specialized approaches to selective attention than the one presented in this paper.

Juergen Schmidhuber 2003-02-21

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