Research map / Architectures and efficiency / RNNs and normalization
Bivariate Beta-LSTM
Kyungwoo Song, JoonHo Jang, Seung jae Shin, Il-Chul Moon
Read the paperCheck your idea against 137,217 papers
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 flexible, correlated Beta-distribution input and forget gates with a structured prior | LSTM · CR, SUBJ, MR, TREC, MPQA, SST | accuracy: 84.66±2.42; 93.25±0.85; 81.59±0.91; 94.80±0.47; 89.66±0.44; 88.94±0.43 | Improved |
| Add bBeta-LSTM(5G+p) gates | SRU(8-layers) · CR, SUBJ, MR, TREC, MPQA, SST | accuracy: 87.00±2.24 -> 87.21±0.78; 93.76±0.61 -> 93.97±0.55; 83.14±1.53 -> 83.51±1.42; 94.52±0.34 -> 94.80±0.47; 90.39±0.55 -> 90 | Improved |
| Add bBeta-LSTM(5G+p) gates | R-Transformer · JSB, Muse, Nottingham, Piano | negative log-likelihood: 8.26±0.03 -> 8.24±0.01; 7.00±0.03 -> 6.19±0.02; 2.24±0.01 -> 2.13±0.08; 7.44±0.03 -> 7.32±0.03 | Improved |
| Add bBeta-LSTM(5G+p) gates | BN-LSTM · sMNIST and pMNIST | error rate: 1.05% -> 0.76%; 4.26±0.50 -> 3.90±0.25 | Improved |
| Add bBeta-LSTM(5G+p) gates | h-detach(0.4), Show & Tell with Resnet152 · MS-COCO | B-4: 30.3 -> 30.8 | Improved |
| Add bBeta-LSTM(5G+p) gates | h-detach(0.4), Show Attend Tell with Resnet152 · MS-COCO | CIDEr: 101.2 -> 103.2 | Improved |
| Use independent Beta-distributed gates instead of correlated gates (ablation) | bBeta-LSTM · CR | accuracy: 84.45±1.87 -> 84.66±2.42 | Hurt |
| Restrict gate correlation to positive values using three Gamma variables (ablation) | bBeta-LSTM · CR | accuracy: 83.63±2.14 -> 84.66±2.42 | Hurt |
| Remove the structured prior (ablation) | bBeta-LSTM(5G+p) · CR | accuracy: 84.48±1.96 -> 84.66±2.42 | Hurt |
| Use bBeta-LSTM(5G+p) instead of LSTM gates | Show & Tell with Resnet152 · MS-COCO | METEOR: 24.8 -> 24.7 | No clear effect |
| Use bBeta-LSTM(5G+p) instead of h-detach(0.4) gates | Show & Tell with Resnet152 and h-detach · MS-COCO | METEOR: 25.2 -> 25.2 | No clear effect |
Closest papers by meaning
- Towards Binary-Valued Gates for Robust LSTM Training (ICML 2018 · similarity 0.88)
- Structured Sparsification of Gated Recurrent Neural Networks (AAAI 2020 · similarity 0.86)
- xLSTM: Extended Long Short-Term Memory (NeurIPS 2024 · similarity 0.85)
- Kernel-Based Approaches for Sequence Modeling: Connections to Neural Methods (NeurIPS 2019 · similarity 0.85)
- Understanding the Differences in Foundation Models: Attention, State Space Models, and Recurrent Neural Networks (NeurIPS 2024 · similarity 0.83)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 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.
