Deep Recurrent Neural Networks for Time Series Prediction

TL;DR

This paper introduces deep recurrent neural networks with extended backpropagation through time for improved modeling and prediction of high-order time series, demonstrated on epileptic seizure data with high accuracy.

Contribution

It presents a novel deep recurrent network architecture with longer backpropagation, analytical error reduction proof, and an efficient DP training method for high-order time series prediction.

Findings

Achieved over 99% average detection rate on seizure data

Demonstrated increased approximation error reduction with added layers

Validated the network's ability to learn temporal structures

Abstract

Ability of deep networks to extract high level features and of recurrent networks to perform time-series inference have been studied. In view of universality of one hidden layer network at approximating functions under weak constraints, the benefit of multiple layers is to enlarge the space of dynamical systems approximated or, given the space, reduce the number of units required for a certain error. Traditionally shallow networks with manually engineered features are used, back-propagation extent is limited to one and attempt to choose a large number of hidden units to satisfy the Markov condition is made. In case of Markov models, it has been shown that many systems need to be modeled as higher order. In the present work, we present deep recurrent networks with longer backpropagation through time extent as a solution to modeling systems that are high order and to predicting ahead. We study epileptic seizure suppression electro-stimulator. Extraction of manually engineered complex features and prediction employing them has not allowed small low-power implementations as, to avoid possibility of surgery, extraction of any features that may be required has to be included. In this solution, a recurrent neural network performs both feature extraction and prediction. We prove analytically that adding hidden layers or increasing backpropagation extent increases the rate of decrease of approximation error. A Dynamic Programming (DP) training procedure employing matrix operations is derived. DP and use of matrix operations makes the procedure efficient particularly when using data-parallel computing. The simulation studies show the geometry of the parameter space, that the network learns the temporal structure, that parameters converge while model output displays same dynamic behavior as the system and greater than .99 Average Detection Rate on all real seizure data tried.