Research map / Architectures and efficiency / RNNs and normalization
Sigsoftmax: Reanalysis of the Softmax Bottleneck
Sekitoshi Kanai, Yasuhiro Fujiwara, Yuki Yamanaka, Shuichi Adachi
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 |
|---|---|---|---|
| Replace softmax with sigsoftmax | 3-layer LSTM · PTB test + dynamic eval. | perplexity: 50.5±0.5 -> 49.2 ± 0.4 | Improved |
| Replace softmax with sigsoftmax | 3-layer LSTM · WT2 test + dynamic eval. | perplexity: 43.3±0.1 -> 42.9 ± 0.1 | Improved |
| Replace MoS with MoSS | 3-layer LSTM mixture model · PTB test + dynamic eval. | perplexity: 48.0±0.1 -> 47.7±0.07 | Improved |
| Replace MoS with MoSS | 3-layer LSTM mixture model · WT2 test + dynamic eval. | perplexity: 40.8±0.03 -> 40.3±0.2 | Improved |
| Use sigmoid-based output instead of softmax (ablation) | 3-layer LSTM · PTB test + dynamic eval. | perplexity: 50.5±0.5 -> 48.9 ± 0.3 | Improved |
| Use sigmoid-based output instead of softmax (ablation) | 3-layer LSTM · WT2 test + dynamic eval. | perplexity: 43.3±0.1 -> 43.5±0.1 | Hurt |
| Use ReLU-based output instead of softmax (ablation) | 3-layer LSTM · PTB test + dynamic eval. | perplexity: 50.5±0.5 -> (2.78±8)$\times 10^5$ | Hurt |
| Use ReLU-based output instead of softmax (ablation) | 3-layer LSTM · WT2 test + dynamic eval. | perplexity: 43.3±0.1 -> (2.30±2)$\times 10^{4}$ | Hurt |
Closest papers by meaning
- On the Softmax Bottleneck of Recurrent Language Models (AAAI 2021 · similarity 0.91)
- Grokking at the Edge of Numerical Stability (ICLR 2025 · similarity 0.84)
- Statistical Perspective of Top-K Sparse Softmax Gating Mixture of Experts (ICLR 2024 · similarity 0.84)
- Dirichlet-Based Prediction Calibration for Learning with Noisy Labels (AAAI 2024 · similarity 0.84)
- Binary Hypothesis Testing for Softmax Models and Leverage Score Models (ICML 2025 · similarity 0.84)
- $\epsilon$-Softmax: Approximating One-Hot Vectors for Mitigating Label Noise (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.
