Adaptive sensory integrating neural network based on a Bayesian estimation method

The authors present a sensory integrating neural network based on a Bayesian method. It is well known that almost all mammals recognize the outer world using several sensors such as eyes and ears. Although mammals basically integrate all sensory inputs, sometimes they ignore a part of the sensory input if the input is very noisy or contradicted from other sensory inputs. Using such adaptive selection strategy, mammals realize robust recognition in any situation. To realize the above in artificial neural networks, we construct a Bayesian method for optimizing the recognition outputs using several sets of forward network and backward network connected to each sensor. In the recognition phase, the system calculates the posterior distribution of the recognition output and the confidence parameter of each sensor, and optimizes the output and the confidence parameter to maximize the posterior distribution function. The repetition of the forward and the backward network calculation realizes the optimization process quickly. The experimental results show that the system yields appropriate recognition results while ignoring the noisy or contradictive sensory inputs by decreasing the corresponding confidence parameter.