Self-supervised learning and recognition by integrating information from several sensors

Almost all of species recognize outer world by integrating information from several sensors such as eyes and ears. Using this strategy, they seems to realize robust recognition in any situations. Furthermore, some researcher found that there are cases that the integration of visual inputs and acoustic inputs plays an important rule for the learning in their early life. From these points of view, we have already proposed a sensory integrating system. The system has several sets of neural networks and sensors. Each neural network receives inputs from corresponding sensor, and recognize the inputs. The output of each neural network is sent to an integrating unit, which integrates outputs from all the neural networks. The integrating unit outputs the recognition result of the system. This paper improves the previous system by introducing a new learning and recognition method based on a Bayesian strategy. The aim of this new system is that 1. Robust recognition in any situation using several sensors. 2. Detection of the correct class of ambiguous sensory inputs, using a Bayesian strategy. 3. Bayesian learning without any supervised signals