Recurrent neural network training with preconditioned stochastic gradient descent

TL;DR

This paper evaluates a preconditioned stochastic gradient descent method for training recurrent neural networks, demonstrating its effectiveness on synthetic and real-world tasks without complex preprocessing.

Contribution

It introduces and tests PSGD, a simple, adaptive preconditioning technique that improves RNN training efficiency and performance without additional tricks.

Findings

PSGD achieves competitive results on synthetic RNN problems.

PSGD performs well on MNIST digit recognition.

No preprocessing or pretraining needed for PSGD.

Abstract

This paper studies the performance of a recently proposed preconditioned stochastic gradient descent (PSGD) algorithm on recurrent neural network (RNN) training. PSGD adaptively estimates a preconditioner to accelerate gradient descent, and is designed to be simple, general and easy to use, as stochastic gradient descent (SGD). RNNs, especially the ones requiring extremely long term memories, are difficult to train. We have tested PSGD on a set of synthetic pathological RNN learning problems and the real world MNIST handwritten digit recognition task. Experimental results suggest that PSGD is able to achieve highly competitive performance without using any trick like preprocessing, pretraining or parameter tweaking.