Analyzing and Exploiting NARX Recurrent Neural Networks for Long-Term Dependencies

TL;DR

This paper introduces MIST RNNs, a NARX architecture that directly connects to distant past states, improving long-term dependency learning, efficiency, and gradient flow compared to LSTM and other NARX models.

Contribution

The paper presents MIST RNNs, a novel NARX architecture that enhances long-term dependency modeling and computational efficiency over existing RNNs.

Findings

MIST RNNs show superior vanishing-gradient properties.

They are more computationally efficient than previous NARX RNNs.

They outperform LSTM and Clockwork RNNs on long-term dependency tasks.

Abstract

Recurrent neural networks (RNNs) have achieved state-of-the-art performance on many diverse tasks, from machine translation to surgical activity recognition, yet training RNNs to capture long-term dependencies remains difficult. To date, the vast majority of successful RNN architectures alleviate this problem using nearly-additive connections between states, as introduced by long short-term memory (LSTM). We take an orthogonal approach and introduce MIST RNNs, a NARX RNN architecture that allows direct connections from the very distant past. We show that MIST RNNs 1) exhibit superior vanishing-gradient properties in comparison to LSTM and previously-proposed NARX RNNs; 2) are far more efficient than previously-proposed NARX RNN architectures, requiring even fewer computations than LSTM; and 3) improve performance substantially over LSTM and Clockwork RNNs on tasks requiring very long-term dependencies.