Wider and Deeper, Cheaper and Faster: Tensorized LSTMs for Sequence Learning

TL;DR

The paper introduces Tensorized LSTMs that expand capacity by widening and deepening networks efficiently through tensor representations and shared parameters, resulting in faster and cheaper sequence learning.

Contribution

It proposes a novel Tensorized LSTM architecture that increases model capacity without additional parameters or runtime, enabling wider and deeper networks for sequence learning.

Findings

Effective on five challenging sequence tasks

Wider networks achieved without extra parameters

Deeper networks with minimal runtime increase

Abstract

Long Short-Term Memory (LSTM) is a popular approach to boosting the ability of Recurrent Neural Networks to store longer term temporal information. The capacity of an LSTM network can be increased by widening and adding layers. However, usually the former introduces additional parameters, while the latter increases the runtime. As an alternative we propose the Tensorized LSTM in which the hidden states are represented by tensors and updated via a cross-layer convolution. By increasing the tensor size, the network can be widened efficiently without additional parameters since the parameters are shared across different locations in the tensor; by delaying the output, the network can be deepened implicitly with little additional runtime since deep computations for each timestep are merged into temporal computations of the sequence. Experiments conducted on five challenging sequence learning tasks show the potential of the proposed model.