Preventing posterior collapse in variational autoencoders for text generation via decoder regularization
Alban Petit, Caio Corro
2021-10-28Text Generation
Abstract
Variational autoencoders trained to minimize the reconstruction error are sensitive to the posterior collapse problem, that is the proposal posterior distribution is always equal to the prior. We propose a novel regularization method based on fraternal dropout to prevent posterior collapse. We evaluate our approach using several metrics and observe improvements in all the tested configurations.
Related Papers
Making Language Model a Hierarchical Classifier and Generator2025-07-17Mitigating Object Hallucinations via Sentence-Level Early Intervention2025-07-16The Devil behind the mask: An emergent safety vulnerability of Diffusion LLMs2025-07-15Seq vs Seq: An Open Suite of Paired Encoders and Decoders2025-07-15Hashed Watermark as a Filter: Defeating Forging and Overwriting Attacks in Weight-based Neural Network Watermarking2025-07-15Exploiting Leaderboards for Large-Scale Distribution of Malicious Models2025-07-11CLI-RAG: A Retrieval-Augmented Framework for Clinically Structured and Context Aware Text Generation with LLMs2025-07-09FIFA: Unified Faithfulness Evaluation Framework for Text-to-Video and Video-to-Text Generation2025-07-09