Twin Networks: Matching the Future for Sequence Generation

TL;DR

This paper introduces Twin Networks, a training technique where a backward RNN guides a forward RNN to better model long-term dependencies in sequence generation tasks, improving performance in speech recognition and captioning.

Contribution

The paper presents a novel training method using a backward RNN to enhance forward RNNs' ability to plan ahead and capture long-term dependencies during sequence generation.

Findings

Achieved 9% relative improvement in speech recognition.

Significant improvement in COCO caption generation.

Backward network used only during training, not inference.

Abstract

We propose a simple technique for encouraging generative RNNs to plan ahead. We train a "backward" recurrent network to generate a given sequence in reverse order, and we encourage states of the forward model to predict cotemporal states of the backward model. The backward network is used only during training, and plays no role during sampling or inference. We hypothesize that our approach eases modeling of long-term dependencies by implicitly forcing the forward states to hold information about the longer-term future (as contained in the backward states). We show empirically that our approach achieves 9% relative improvement for a speech recognition task, and achieves significant improvement on a COCO caption generation task.