Representing dynamically: An active process for describing sequential data

We propose an unsupervised method for building dynamic representations of sequential data, particularly of observed interactions. The method simultaneously acquires representations of input data and its dynamics. It is based on a hierarchical generative model composed of two levels. In the first level, a model learns representations to generate observed data. In the second level, representational states encode the dynamics of the lower one. The model is designed as a Bayesian network with switching variables represented in the higher level, and which generates transition models. The method actively explores the latent space guided by its knowledge and the uncertainty about it. That is achieved by updating the latent variables from prediction error signals backpropagated to the latent space. So, no encoder or inference models are used since the generators also serve as their inverse transformations. The method is evaluated in two scenarios, with static images and with videos. The results show that the adaptation over time leads to better performance than with similar architectures without temporal dependencies, e.g., variational autoencoders. With videos, it is shown that the system extracts the dynamics of the data in states that highly correlate with the ground truth of the actions observed.