No abstract

The combination of reinforcement learning with deep learning is a promising approach to tackle important sequential decision-making problems that are currently intractable. One obstacle to overcome is the amount of data needed by learning systems of this type. In this article, we propose to address this issue through a divide-and-conquer approach. We argue that complex decision problems can be naturally decomposed into multiple tasks that unfold in sequence or in parallel. By associating each task with a reward function, this problem decomposition can be seamlessly accommodated within the standard reinforcement-learning formalism. The specific way we do so is through a generalization of two fundamental operations in reinforcement learning: policy improvement and policy evaluation. The generalized version of these operations allow one to leverage the solution of some tasks to speed up the solution of others. If the reward function of a task can be well approximated as a linear combination of the reward functions of tasks previously solved, we can reduce a reinforcement-learning problem to a simpler linear regression. When this is not the case, the agent can still exploit the task solutions by using them to interact with and learn about the environment. Both strategies considerably reduce the amount of data needed to solve a reinforcement-learning problem.

In this paper, we present a tracking framework for capturing articulated human motions in real-time, without the need for attaching markers onto the subject's body. This is achieved by first obtaining a low dimensional representation of the training motion data, using a nonlinear dimensionality reduction technique called back-constrained GPLVM. A prior dynamics model is then learnt from this low dimensional representation by partitioning the motion sequences into elementary movements using an unsupervised EM clustering algorithm. The temporal dependencies between these elementary movements are efficiently captured by a Variable Length Markov Model. The learnt dynamics model is used to bias the propagation of candidate pose feature vectors in the low dimensional space. By combining this with an efficient volumetric reconstruction algorithm, our framework can quickly evaluate each candidate pose against image evidence captured from multiple views. We present results that show our system can accurately track complex structured activities such as ballet dancing in real-time.

We present a novel approach to speech-driven facial animation using a non-parametric switching state space model based on Gaussian processes. The model is an extension of the shared Gaussian process dynamical model, augmented with switching states. Audio and visual data from a talking head corpus are jointly modelled using the proposed method. The switching states are found using variable length Markov models trained on labelled phonetic data. We also propose a synthesis technique that takes into account both previous and future phonetic context, thus accounting for coarticulatory effects in speech.

No abstract