A Recurrent Anticipatory Self-Organized Map for Robust Time Series Prediction A Recurrent Anticipatory Self-Organized Map for Robust Time Series Prediction

In robotics applications, we often have noisy data that have temporal constraints due to the real world, such as sensory sequences generated by the motion of a robot through the environment. To recover the underlying structure of this noisy stream of data, we can do clustering with a self-organized map (SOM). We can view the SOM as a generative model, and it is seen to be doing maximum-likelihood estimation in that case. It has been shown that when the SOM is working with a temporal stream of data, it is possible to convert it to a maximum a posteriori estimator by adding lateral weights between the SOM nodes. However, this method has the drawback that it assumes that the temporal data are first order Markovian. In this paper we show how to overcome this limitation. We use a simple recurrent network (SRN) to predict the priors for the SOM. Since the SRN is capable of learning sequences in the data, this overcomes the main limitation of the previous model. We also bring out the connection between our model (the RecAntSOM) and the standard predictor-corrector framework of filtering. We do some simple robot experiments to show the usefulness of our model.