Capacity and Trainability in Recurrent Neural Networks

Abstract

Two potential bottlenecks on the expressiveness of recurrent neural networks (RNNs) are their ability to store information about the task in their parameters, and to store information about the input history in their units. We show experimentally that all common RNN architectures achieve nearly the same per-task and per-unit capacity bounds with careful training, for a variety of tasks and stacking depths. They can store an amount of task information which is linear in the number of parameters, and is approximately 5 bits per parameter. They can additionally store approximately one real number from their input history per hidden unit. We further find that for several tasks it is the per-task parameter capacity bound that determines performance. These results suggest that many previous results comparing RNN architectures are driven primarily by differences in training effectiveness, rather than differences in capacity. Supporting this observation, we compare training difficulty for several architectures, and show that vanilla RNNs are far more difficult to train, yet have slightly higher capacity. Finally, we propose two novel RNN architectures, one of which is easier to train than the LSTM or GRU for deeply stacked architectures.

Published as a conference paper at ICLR 2017

Cited by in corpus (29)

• Recurrent Neural Networks for Time Series Forecasting: Current Status and Future Directions
• Efficient Neural Architecture Search via Parameter Sharing
• Underspecification Presents Challenges for Credibility in Modern Machine Learning
• Deep Learning Scaling is Predictable, Empirically
• Contextual Hybrid Session-based News Recommendation with Recurrent Neural Networks
• Overcoming the vanishing gradient problem in plain recurrent networks
• Recurrent Neural Networks for Semantic Instance Segmentation
• Dual Rectified Linear Units (DReLUs): A Replacement for Tanh Activation Functions in Quasi-Recurrent Neural Networks
• Universality and individuality in neural dynamics across large populations of recurrent networks
• Deep Independently Recurrent Neural Network (IndRNN)
• A Taxonomy for Neural Memory Networks
• Is it enough to optimize CNN architectures on ImageNet?
• AdaFrame: Adaptive Frame Selection for Fast Video Recognition
• TraDE: Transformers for Density Estimation
• Demystifying Deep Learning in Predictive Spatio-Temporal Analytics: An Information-Theoretic Framework
• Understanding How Encoder-Decoder Architectures Attend
• On Training Recurrent Networks with Truncated Backpropagation Through Time in Speech Recognition
• Thalamocortical motor circuit insights for more robust hierarchical control of complex sequences
• On Predictive Information in RNNs
• The geometry of integration in text classification RNNs
• RNNs Evolving on an Equilibrium Manifold: A Panacea for Vanishing and Exploding Gradients?
• Towards an efficient deep learning model for musical onset detection
• Tree-Augmented Cross-Modal Encoding for Complex-Query Video Retrieval
• Rational Recurrences
• Persistence pays off: Paying Attention to What the LSTM Gating Mechanism Persists
• Quantitative Analysis of Image Classification Techniques for Memory-Constrained Devices
• An Empirical Exploration of Deep Recurrent Connections and Memory Cells Using Neuro-Evolution
• Unfolding recurrence by Green's functions for optimized reservoir computing
• Learning Longer-term Dependencies via Grouped Distributor Unit