Feedforward Sequential Memory Neural Networks without Recurrent Feedback

TL;DR

This paper presents FSMN, a novel feedforward neural network architecture with learnable memory blocks that effectively capture long-term dependencies in language modeling tasks without recurrent feedback.

Contribution

The paper introduces FSMN, a new memory neural network structure that learns long-term dependencies without recurrence, outperforming traditional FNN and RNN language models.

Findings

FSMN effectively captures long-term dependencies.

FSMN-based language models outperform FNN and RNN models.

Experimental results demonstrate significant performance improvements.

Abstract

We introduce a new structure for memory neural networks, called feedforward sequential memory networks (FSMN), which can learn long-term dependency without using recurrent feedback. The proposed FSMN is a standard feedforward neural networks equipped with learnable sequential memory blocks in the hidden layers. In this work, we have applied FSMN to several language modeling (LM) tasks. Experimental results have shown that the memory blocks in FSMN can learn effective representations of long history. Experiments have shown that FSMN based language models can significantly outperform not only feedforward neural network (FNN) based LMs but also the popular recurrent neural network (RNN) LMs.