Research map / Architectures and efficiency / RNNs and normalization
Trainable Decoding of Sets of Sequences for Neural Sequence Models
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 |
|---|---|---|---|
| Learned submodular set decoder trained with CE, REINFORCE and end-to-end finetuning (∇BS-CE-RE-EE) vs. beam search | LSTM captioner with K=5 outputs · COCO | Faccuracy (K=5): 0.3063 vs 0.2842 | Helped |
| Adding REINFORCE training after CE (∇BS-CE-RE) over CE-only training (ablation) | ∇BS decoder · COCO | Faccuracy (K=5): 0.3056 vs 0.3015 | Helped |
| Diverse beam search vs. beam search baseline | LSTM captioner · COCO | distinct 4-grams: 68.18 vs 20.04 | Helped |
| ∇BS training (CE-RE-EE) relative to diverse beam search in diversity | ∇BS decoder · COCO | distinct 4-grams: 35.84 vs 68.18 | Hurt |
| Adding end-to-end finetuning of LSTM after CE-RE (ablation) | ∇BS-CE-RE · Flickr-8k | Faccuracy (K=5): 0.2712 vs 0.2707 | Helped |
| Adding end-to-end finetuning of LSTM after CE (∇BS-CE-EE) over CE-only (ablation) | ∇BS-CE · COCO | Faccuracy (K=5): 0.3011 vs 0.3015 | No clear effect |
| MIXER training vs. CE-RE-EE on Flickr-8k (ablation) | ∇BS decoder · Flickr-8k | Faccuracy (K=5): 0.2697 vs 0.2712 | Mixed |
| Trainable decoder with REINFORCE (SCST-style) vs beam search trained with same metric on Flickr-8k | LSTM captioner · Flickr-8k | Faccuracy (K=5): 0.2643 vs 0.2510 | Helped |
Closest papers by meaning
- Non-autoregressive Sequence-to-Sequence Vision-Language Models (CVPR 2024 · similarity 0.87)
- Latent Sequence Decompositions (ICLR 2017 · similarity 0.87)
- Empirical Analysis of Beam Search Performance Degradation in Neural Sequence Models (ICML 2019 · similarity 0.86)
- Efficiently Modeling Long Sequences with Structured State Spaces (ICLR 2022 · similarity 0.86)
- Convolutional Sequence to Sequence Learning (ICML 2017 · similarity 0.86)
- Progressive Inference: Explaining Decoder-Only Sequence Classification Models Using Intermediate Predictions (ICML 2024 · similarity 0.85)
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.