Research map / Architectures and efficiency / RNNs and normalization
Bidirectional Beam Search: Forward-Backward Inference in Neural Sequence Models for Fill-In-The-Blank Image Captioning
Check 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 bidirectional beam search instead of the listed baselines | LSTM fill-in-the-blank captioning · COCO, known blank length | CIDEr, Bleu-4, Meteor: r=25%: CIDEr 6.98 -> 7.12, Bleu-4 0.709 -> 0.720, Meteor 0.510 -> 0.517; r=50%: CIDEr 4.15 -> 4.26, Bleu-4 0.398 -> 0.40 | Helped |
| Use bidirectional beam search instead of the listed baselines | LSTM fill-in-the-blank captioning · COCO, unknown blank length | CIDEr, Bleu-4, Meteor: r=25%: CIDEr 5.725 -> 5.935, Bleu-4 0.589 -> 0.614, Meteor 0.452 -> 0.460; r=50%: CIDEr 4.380 -> 4.40, Bleu-4 0.453 -> 0 | Mixed |
| Use bidirectional beam search rather than generation baselines | CNN+LSTM fill-in-the-blank generation · Visual Madlibs type 7, known length | Bleu-1, Bleu-2: 0.460 -> 0.470; 0.284 -> 0.389 | Mixed |
| Use bidirectional beam search rather than generation baselines | CNN+LSTM fill-in-the-blank generation · Visual Madlibs type 12, known length | Bleu-1, Bleu-2: 0.347 -> 0.353; 0.214 -> 0.216 | Helped |
| Use bidirectional beam search instead of the strongest generation baseline | CNN+LSTM fill-in-the-blank generation · Visual Madlibs type 7, unknown length | Bleu-1, Bleu-2: 0.348 -> 0.351; 0.203 -> 0.197 | Mixed |
| Use bidirectional beam search instead of the strongest generation baseline | CNN+LSTM fill-in-the-blank generation · Visual Madlibs type 12, unknown length | Bleu-1, Bleu-2: 0.309 -> 0.31; 0.186 -> 0.190 | Helped |
| Use backward-only rather than forward-only standard beam search | LSTM caption model · COCO fill-in-the-blank, known length | CIDEr, Bleu-4, Meteor: URNN-f -> URNN-b: r=25% CIDEr 6.54 -> 6.58, Bleu-4 0.661 -> 0.668, Meteor 0.488 -> 0.491; r=50% CIDEr 3.744 -> 3.931, Bl | Helped |
| Add bidirectional context to decoding rather than using unidirectional beam search | Fill-in-the-blank sequence generation · COCO and Visual Madlibs | sentence-level BLEU and caption metrics | Helped |
| Alternate forward and backward beam updates to approximate the full joint | BiRNN-BiBS · Fill-in-the-blank image captioning | joint log probability: -2.47 to -2.11 | Helped |
Closest papers by meaning
- Empirical Analysis of Beam Search Performance Degradation in Neural Sequence Models (ICML 2019 · similarity 0.90)
- Diverse Beam Search for Improved Description of Complex Scenes (AAAI 2018 · similarity 0.89)
- A Continuous Relaxation of Beam Search for End-to-End Training of Neural Sequence Models (AAAI 2018 · similarity 0.86)
- Trainable Decoding of Sets of Sequences for Neural Sequence Models (ICML 2019 · similarity 0.86)
- Linear Attention for Efficient Bidirectional Sequence Modeling (NeurIPS 2025 · similarity 0.84)
- Non-autoregressive Sequence-to-Sequence Vision-Language Models (CVPR 2024 · 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.