Research map / Architectures and efficiency / Quantization
HitNet: Hybrid Ternary Recurrent Neural Network
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 TTQ for weights and BTQ for activations (hybrid) instead of TTQ for both, without sloping factor (ablation) | ternary LSTM · PTB | PPW: 125.5 -> 120.1 | Helped |
| Add sloping factor lambda=0.4 to activation functions of HitNet (ablation) | HitNet LSTM · PTB | PPW: 120.1 -> 110.3 | Helped |
| Use too small sloping factor lambda=0.1 (ablation) | HitNet LSTM · PTB | PPW: 136.4 vs 120.1 at lambda=1 | Hurt |
| Small sloping factor lambda=0.3-0.5 on full-precision LSTM (ablation) | Full-precision LSTM · PTB | PPW: 97.2 -> 95.8 | Helped |
| Quantize both weights and activations with 2-bit uniform quantization vs weights only | LSTM · PTB | PPW: 103.3 -> 638.9 | Hurt |
| Quantize activations with TTQ rather than weights only | LSTM · PTB | PPW: 108.7 -> 125.5 | Hurt |
| HitNet vs prior 2-bit quantization (Zhou et al.) | LSTM · PTB | PPW: 126 -> 110.3 | Helped |
| HitNet vs prior 2-bit quantization (Zhou et al.) | GRU · PTB | PPW: 142 -> 113.5 | Helped |
Closest papers by meaning
- Alternating Multi-bit Quantization for Recurrent Neural Networks (ICLR 2018 · similarity 0.89)
- Trained Ternary Quantization (ICLR 2017 · similarity 0.88)
- Learning Recurrent Binary/Ternary Weights (ICLR 2019 · similarity 0.88)
- RTN: Reparameterized Ternary Network (AAAI 2020 · similarity 0.87)
- M$^2$RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling (NeurIPS 2026 · similarity 0.86)
- HADAMRNN: BINARY AND SPARSE TERNARY ORTHOGONAL RNNS (ICLR 2025 · similarity 0.86)
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.