Multitask Pattern Recognition for Vision-Based Autonomous Robots

This paper uses Multitask Pattern Recognition (MTPR) to improve the accuracy and robustness of neural net based vision systems. MTPR trains neural nets on a set of auxiliary recognition problems at the same time the net is trained on the main recognition task. The predictions made for the auxiliary tasks are not used, but the internal features learned by the net for them improve performance on the main recognition task. The auxiliary tasks allow us to focus attention towards features that learning would otherwise ignore. MTPR is broadly applicable. It improves performance on a simulated ALVINN domain 10%-30%.