On the use of intrinsic motivation for visual saliency learning

The use of intrinsic motivation for the task of learning sensori-motor properties has received a lot of attention over the last few years, but only little work has been provided toward using intrinsic motivation for the task of learning visual signals. In this paper, we propose to apply the main ideas of the Intelligent Adaptive Curiosity (IAC) for the task of visual saliency learning. We here present RL-IAC, an adapted version of IAC that uses reinforcement learning to deal with time consuming displacements while actively learning saliency based on local learning progress. We also introduce the use of a backward evaluation to deal with a learner that is shared between several regions. We demonstrate the good performance of RL-IAC compared to other exploration techniques, and we discuss the performance of other intrinsic motivation sources instead of learning progress in our problem.