Research map / Architectures and efficiency / RNNs and normalization
Imputer: Sequence Modelling via Imputation and Dynamic Programming
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Dynamic Programming instead of Imitation Learning (ablation) | Imputer · WSJ | WER: 16.5 -> 12.7 | Helped |
| use Dynamic Programming instead of Imitation Learning (ablation) | Imputer · LibriSpeech test-other | WER: 14.6 -> 11.1 | Helped |
| use Dynamic Programming instead of Imitation Learning (ablation) | Imputer · LibriSpeech test-clean | WER: 5.5 -> 4.0 | Helped |
| use Dynamic Programming instead of Imitation Learning (ablation) | Imputer · WSJ | CER: 6.2 -> 4.9 | Helped |
| use CTC-based alignment roll-in rather than decoding Imputer samples during training (ablation) | Imputer · LibriSpeech | WER | No clear effect |
| add noise by shifting alignments up to 1 frame left or right (ablation) | Imputer alignment roll-in · LibriSpeech | WER: <0.2 | Helped |
| use Bernoulli masking (ablation) | Imputer · LibriSpeech dev-other | WER: 11.2 | No clear effect |
| use Uniform masking (ablation) | Imputer · LibriSpeech dev-other | WER: 11.4 | No clear effect |
| replace block inference with greedy top-k decoding (ablation) | Imputer · LibriSpeech dev-other | WER: 11.2 -> 11.6 | Hurt |
| train and infer with smaller block sizes than B=8 (ablation) | Imputer · LibriSpeech dev-other | WER | Hurt |
| train and infer with larger block sizes than B=8 (ablation) | Imputer · LibriSpeech dev-other | WER | Hurt |
| use Right Most Last decoding with a larger inference block size (ablation) | Imputer trained with block size B=8 · LibriSpeech dev-other | WER | Helped |
Closest papers by meaning
- Optimal Completion Distillation for Sequence Learning (ICLR 2019 · similarity 0.87)
- Sequence to Multi-Sequence Learning via Conditional Chain Mapping for Mixture Signals (NeurIPS 2020 · similarity 0.87)
- Trainable Decoding of Sets of Sequences for Neural Sequence Models (ICML 2019 · similarity 0.85)
- Latent Sequence Decompositions (ICLR 2017 · similarity 0.85)
- Align With Purpose: Optimize Desired Properties in CTC Models with a General Plug-and-Play Framework (ICLR 2024 · similarity 0.83)
- UniWav: Towards Unified Pre-training for Speech Representation Learning and Generation (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.