Research map / Architectures and efficiency / RNNs and normalization
Learning Hierarchical Structures with Differentiable Nondeterministic Stacks
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 |
|---|---|---|---|
| Use unnormalized transition weights and include PDA states in the stack reading | NS-RNN · Unmarked reversal | per-symbol cross-entropy | Helped |
| Add unnormalized transition weights to the NS-RNN (ablation) | NS-RNN · Hardest CFL | per-symbol cross-entropy | Helped |
| Add PDA states to the stack reading (ablation) | NS-RNN · Padded reversal | per-symbol cross-entropy | Helped |
| Add unnormalized transition weights to the NS-RNN (ablation) | NS-RNN · Padded reversal | training convergence | Helped |
| Add PDA states to the stack reading (ablation) | NS-RNN · Unmarked reversal | per-symbol cross-entropy | Helped |
| Add unnormalized transition weights (ablation) | NS-RNN · Unmarked reversal | per-symbol cross-entropy | Helped |
| Use unnormalized transition weights and PDA-state readings | NS-RNN · Hardest CFL | per-symbol cross-entropy | Helped |
| Add unnormalized transition weights (ablation) | NS-RNN · Marked reversal | learning speed | Helped |
| Replace an LSTM with a superposition stack (JM) | Penn Treebank language model · Penn Treebank test | perplexity: 120.95 -> 115.35 | Helped |
| Replace an LSTM with JM pushing the hidden state | Penn Treebank language model · Penn Treebank test | SG score: 0.433 -> 0.387 | Hurt |
| Apply RNS modifications rather than normalized NS actions (ablation) | NS-RNN, |Q| = 1, |Γ| = 2 · Penn Treebank test | perplexity: 122.62 -> 117.56 | Helped |
| Apply RNS modifications rather than normalized NS actions (ablation) | NS-RNN, |Q| = 1, |Γ| = 2 · Penn Treebank test | SG score: 0.414 -> 0.435 | No clear effect |
Closest papers by meaning
- The Surprising Computational Power of Nondeterministic Stack RNNs (ICLR 2023 · similarity 0.91)
- Memory Architectures in Recurrent Neural Network Language Models (ICLR 2018 · similarity 0.90)
- Modeling Hierarchical Structures with Continuous Recursive Neural Networks (ICML 2021 · similarity 0.89)
- Hierarchically Gated Recurrent Neural Network for Sequence Modeling (NeurIPS 2023 · similarity 0.89)
- State-Regularized Recurrent Neural Networks (ICML 2019 · similarity 0.89)
- M$^2$RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling (NeurIPS 2026 · similarity 0.88)
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.