Visual reinforcement learning for object recognition in robotics

In recent years, unmanned systems reached a maturity level that resulted in the adoption of robotic system for both civil and military usages. As the amount of unmanned systems increases, the amount of visual information that is being sent to the end user increases, and the need to develop the unmanned systems ability to autonomously alert the user on predefined events is being emphasized. This ability is mainly based on their capability to detect and classify the objects in their field of operation. Recent studies shows that the mechanism behind the recognition process in the human brain includes continuous generation of predictions based on prior knowledge about the world. These predictions enable rapid generation of hypothesis that biases the outcome of the recognition process in situation of uncertainty. In addition, the visual system continuously updating its knowledge about the world based on visual feedback. In our work, we formed an architecture that is based on the concepts behind these top-down prediction and learning processes of the human visual system, together with state of the art bottom-up object recognition models, e.g. deep CNNs. We show that imitating these top-down reinforcement learning mechanisms, together with the state of the art bottom-up feed-forward algorithm, creates an accurate and continuously improving target recognition model.