Acquisition of context-based active word recognition by q-learning using a recurrent neural network

Abstract

In the real world, where there is a large amount of information, humans recognize an object efficiently by moving their sensors, and if it is supported by context information, a better result could be produced. In this paper, the emergence of sensor motion and a contextbased recognition function are expected. The sensor-equipped recognition learning system has a very simple and general architecture that is consisted of one recurrent neural network and trained by reinforcement learning. The proposed learning system learns to move a visual sensor intentionally and to classify a word from the series of partial information simultaneously only based on the reward and punishment generated from the recognition result. After learning, it was verified that the contextbased word recognition could be achieved. All words were correctly recognized at the appropriate timing by actively moving the sensors not depending on the initial sensor location.