A model for learning to segment temporal sequences, utilizing a mixture of RNN experts together with adaptive variance

TL;DR

This paper introduces a new learning method for a mixture of RNN experts that can generate complex sequences by dynamically switching experts, with improved variance control and successful application to robotic sensory-motor data.

Contribution

It presents a novel maximum likelihood-based learning algorithm that incorporates adaptive variance for each expert, enabling better sequence segmentation and generation.

Findings

Successfully learned Markov chain switching among 9 Lissajous curves

Outperformed conventional methods in generalization capability

Applied to sensory-motor flow learning in a humanoid robot

Abstract

This paper proposes a novel learning method for a mixture of recurrent neural network (RNN) experts model, which can acquire the ability to generate desired sequences by dynamically switching between experts. Our method is based on maximum likelihood estimation, using a gradient descent algorithm. This approach is similar to that used in conventional methods; however, we modify the likelihood function by adding a mechanism to alter the variance for each expert. The proposed method is demonstrated to successfully learn Markov chain switching among a set of 9 Lissajous curves, for which the conventional method fails. The learning performance, analyzed in terms of the generalization capability, of the proposed method is also shown to be superior to that of the conventional method. With the addition of a gating network, the proposed method is successfully applied to the learning of sensory-motor flows for a small humanoid robot as a realistic problem of time series prediction and generation.