Scenes With Multiple Objects

Scenes with multiple objects or objects with rich internal details require either recurrent connections in both $M$ and $C$ or some other mechanism for escaping certain cases of local minima. Local minima can be caused by parts of the pixel plane that look similar to the target input, while the nearby environment does not. In such cases the relevant external feedback through the environment becomes non-Markovian. For such situations, additional experiments (not reported here) with a recurrent $M$ and a recurrent $C$ were conducted. It turned out that internal feedback within $M$ and $C$ sometimes can lead to success in cases where the simple approach fails [1]. However, there is an approach which in the long run might prove to be even more promising: The adaptive on-line generation of appropriate sub-goals. Some first work in this direction already has been done [13]. By using the above-mentioned concept of goal-defining input units with time-invariant activations, we intend to apply adaptive sub-goal generators to the problems of `local minima' that can arise during the target detection process.

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