Cyclical Annealing Schedule: A Simple Approach to Mitigating KL Vanishing

Abstract

Variational autoencoders (VAEs) with an auto-regressive decoder have been applied for many natural language processing (NLP) tasks. The VAE objective consists of two terms, (i) reconstruction and (ii) KL regularization, balanced by a weighting hyper-parameter β. One notorious training difficulty is that the KL term tends to vanish. In this paper we study scheduling schemes for β, and show that KL vanishing is caused by the lack of good latent codes in training the decoder at the beginning of optimization. To remedy this, we propose a cyclical annealing schedule, which repeats the process of increasing βmultiple times. This new procedure allows the progressive learning of more meaningful latent codes, by leveraging the informative representations of previous cycles as warm re-starts. The effectiveness of cyclical annealing is validated on a broad range of NLP tasks, including language modeling, dialog response generation and unsupervised language pre-training.

arXiv commentPublished in NAACL 2019; The first two authors contribute equally; Code: https://github.com/haofuml/cyclical_annealing

References in corpus (7)

• How NOT To Evaluate Your Dialogue System: An Empirical Study of Unsupervised Evaluation Metrics for Dialogue Response Generation
• InfoVAE: Information Maximizing Variational Autoencoders
• Snapshot Ensembles: Train 1, get M for free
• Cyclical Stochastic Gradient MCMC for Bayesian Deep Learning
• Amortized Inference Regularization
• Avoiding Latent Variable Collapse With Generative Skip Models
• Stochastic WaveNet: A Generative Latent Variable Model for Sequential Data

Cited by in corpus (39)

• Gravitational-wave parameter estimation with autoregressive neural network flows
• Cyclical Stochastic Gradient MCMC for Bayesian Deep Learning
• Generative Chemical Transformer: Neural Machine Learning of Molecular Geometric Structures from Chemical Language via Attention
• Conditional Flow Variational Autoencoders for Structured Sequence Prediction
• ControlVAE: Controllable Variational Autoencoder
• Generative Capacity of Probabilistic Protein Sequence Models
• Semi-Supervised Learning of Bearing Anomaly Detection via Deep Variational Autoencoders
• Temporally coherent backmapping of molecular trajectories from coarse-grained to atomistic resolution
• Merger-Ringdown Consistency: A New Test of Strong Gravity using Deep Learning
• Dynamics-aware Embeddings
• Bayesian Graph Neural Networks with Adaptive Connection Sampling
• PTP: Parallelized Tracking and Prediction with Graph Neural Networks and Diversity Sampling
• Supplementing Recurrent Neural Networks with Annealing to Solve Combinatorial Optimization Problems
• DLow: Diversifying Latent Flows for Diverse Human Motion Prediction
• Optimus: Organizing Sentences via Pre-trained Modeling of a Latent Space
• Explainable Representation Learning of Small Quantum States
• Annealed Stein Variational Gradient Descent
• Conditional Noise Deep Learning for Parameter Estimation of Gravitational Wave Events
• Jointly Optimizing Diversity and Relevance in Neural Response Generation
• Bayesian Neural Networks at Scale: A Performance Analysis and Pruning Study
• On Stein Variational Neural Network Ensembles
• Single-partition adaptive Q-learning
• An Adversarial Transfer Network for Knowledge Representation Learning
• A Batch Normalized Inference Network Keeps the KL Vanishing Away
• Universal Design Methodology for Printable Microstructural Materials via a New Deep Generative Learning Model: Application to a Piezocomposite
• Unsupervised dynamic modeling of medical image transformation
• Controllable and Diverse Text Generation in E-commerce
• Autoencoders as Tools for Program Synthesis
• Acceleration via Fractal Learning Rate Schedules
• POSGen: Personalized Opening Sentence Generation for Online Insurance Sales
• ControlVAE: Tuning, Analytical Properties, and Performance Analysis
• Russian Natural Language Generation: Creation of a Language Modelling Dataset and Evaluation with Modern Neural Architectures
• Variational Knowledge Distillation for Disease Classification in Chest X-Rays
• Mathematical Word Problem Generation from Commonsense Knowledge Graph and Equations
• Uncertainty-Aware Learning for Improvements in Image Quality of the Canada-France-Hawaii Telescope
• Latte-Mix: Measuring Sentence Semantic Similarity with Latent Categorical Mixtures
• Gradient Boosted Normalizing Flows
• Semi-Implicit Stochastic Recurrent Neural Networks
• Learning from Multiple Sources for Data-to-Text and Text-to-Data