Research map / Label-efficient and robust learning / Knowledge distillation
EM-Network: Oracle Guided Self-distillation for Sequence Learning
Ji Won Yoon, SungHwan Ahn, Hyeonseung Lee, Minchan Kim, Seok Min Kim, Nam Soo Kim
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Add target-derived oracle guidance and self-distillation | Conformer-CTC Small · LibriSpeech test-clean, w/o LM | WER (%): 4.87 -> 4.29 | Improved |
| Add target-derived oracle guidance and self-distillation | Conformer-CTC Small · LibriSpeech test-other, w/o LM | WER (%): 12.05 -> 10.81 | Improved |
| Add target-derived oracle guidance and self-distillation | data2vec Base · LibriSpeech test-clean, greedy decoding | WER (%): 2.78 -> 2.66 | Improved |
| Add target-derived oracle guidance and self-distillation | Transformer · IWSLT'14 En-De test | BLEU: 29.43 -> 29.94 | Improved |
| Add target-derived oracle guidance and self-distillation | Transformer · IWSLT'14 De-En test | BLEU: 35.99 -> 36.53 | Improved |
| Add target-derived oracle guidance and self-distillation | BiBERT · WMT'14 newstest2014 En-De | BLEU: 31.26 -> 31.30 | Improved |
| Add target-derived oracle guidance and self-distillation | BiBERT · WMT'14 newstest2014 De-En | BLEU: 34.94 -> 35.40 | Improved |
| Train with online self-distillation rather than offline distillation (ablation) | EM-Network with Conformer-CTC Small · LibriSpeech test-clean | WER (%): 4.35 -> 4.29 | Improved |
| Train with online self-distillation rather than offline distillation (ablation) | EM-Network with Conformer-CTC Small · LibriSpeech test-other | WER (%): 10.97 -> 10.81 | Improved |
| Add EM-Network self-distillation to the hybrid CTC/attention model | Hybrid CTC/attention Transformer · LibriSpeech test-clean with LM | WER (%): 4.1 -> 3.6 | Improved |
| Add EM-Network self-distillation to the hybrid CTC/attention model | Hybrid CTC/attention Transformer · LibriSpeech test-other with LM | WER (%): 9.4 -> 8.8 | Improved |
| Use EM-Network fine-tuning rather than data2vec fine-tuning | data2vec · LibriSpeech test-other, 100h fine-tuning | WER (%): 9.8 -> 9.8 | No clear effect |
Closest papers by meaning
- Optimal Completion Distillation for Sequence Learning (ICLR 2019 · similarity 0.85)
- The Missing Alignment Link of In-context Learning on Sequences (ICML 2025 · similarity 0.84)
- Towards Understanding Self-Pretraining for Sequence Classification (NeurIPS 2026 · similarity 0.83)
- Heuristic-free Knowledge Distillation for Streaming ASR via Multi-modal Training (AAAI 2025 · similarity 0.83)
- Self-Distillation Regularized Connectionist Temporal Classification Loss for Text Recognition: A Simple Yet Effective Approach (AAAI 2024 · similarity 0.83)
- DistiLLM: Towards Streamlined Distillation for Large Language Models (ICML 2024 · similarity 0.82)
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.
