LSTM: A Search Space Odyssey

TLDR

Since 1995, many LSTM variants have been proposed and are now state‑of‑the‑art for diverse machine‑learning tasks, prompting renewed interest in their computational components. This study conducts the first large‑scale comparison of eight LSTM variants across speech, handwriting, and polyphonic music tasks. We optimized each variant’s hyperparameters with random search and quantified their importance using fANOVA over 5,400 runs, roughly 15 years of CPU time. None of the variants outperformed the standard LSTM; the forget gate and output activation proved most critical, while hyperparameters were largely independent, yielding practical tuning guidelines.

Abstract

Several variants of the Long Short-Term Memory (LSTM) architecture for recurrent neural networks have been proposed since its inception in 1995. In recent years, these networks have become the state-of-the-art models for a variety of machine learning problems. This has led to a renewed interest in understanding the role and utility of various computational components of typical LSTM variants. In this paper, we present the first large-scale analysis of eight LSTM variants on three representative tasks: speech recognition, handwriting recognition, and polyphonic music modeling. The hyperparameters of all LSTM variants for each task were optimized separately using random search, and their importance was assessed using the powerful fANOVA framework. In total, we summarize the results of 5400 experimental runs ($\approx 15$ years of CPU time), which makes our study the largest of its kind on LSTM networks. Our results show that none of the variants can improve upon the standard LSTM architecture significantly, and demonstrate the forget gate and the output activation function to be its most critical components. We further observe that the studied hyperparameters are virtually independent and derive guidelines for their efficient adjustment.