Automatic Segmentation of Continuous Time-Series Data Based on Prediction Error Using Deep Predictive Learning

In deep learning, training with multiple modalities generally improves the learning performance. In robot motion learning, human knowledge information can be regarded as additional modalities and aids robots to perform their tasks. Previous studies have shown that annotation of action segments improves the generalizability of motion learning. However, manual segmentation causes inconsistencies of segmentation because of its subjectivity, Additionally manual segmentation is time-consuming for large-scale datasets. Automatic segmentation by generative models, such as hidden Markov models and predictive learning models, is an alternative approach to find the action segments and may be replaced with human knowledge information. In this study, we expand a predictive learning model using recurrent neural networks to automatic segmentation of continuous time-series data as a robot behavior. We assume that robot behavior can be described as sequences of complex time series composed of multiple simple patterns. As such, we utilize a learning model that can segment data into patterns through predictive learning. To do so, the continuous data is first segmented equally, and next the predictive learning model autonomously shifts the segmentation boundaries to minimize its prediction error. We conducted a numerical experiment that simulates a time series of sensor information. Comparing with the hidden Markov model, the results showed that our predictive learning model were able to estimate more suitable segmenting points.