Research map / Architectures and efficiency / RNNs and normalization
Neural Power Units
Niklas Heim, Tomas Pevny, Vasek Smidl
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 |
|---|---|---|---|
| Use NPU rather than NALU | two-layer arithmetic models · Simple arithmetic task (+, ×, ÷, √) | testing error: NALU: 2.69 ± 0.22, 4.55 ± 0.2, 3.33 ± 0.18, 0.034 ± 0.006; NPU: 0.2 ± 0.11, 0.37 ± 0.23, 0.23 ± 0.13, 0.031 ± 0.025 | Improved |
| Remove imaginary parameters from NPU (ablation) | NPU · Simple arithmetic task (+, ×, ÷, √) | testing error: NPU: 0.2 ± 0.11, 0.37 ± 0.23, 0.23 ± 0.13, 0.031 ± 0.025; RealNPU: 0.08 ± 0.021, 0.066 ± 0.026, 0.085 ± 0.038, 0.004 ± 0 | Improved |
| Use NMU rather than NPU | arithmetic models · Simple arithmetic task (×) | testing error: 0.37 ± 0.23 -> 0.005 ± 0.004 | Improved |
| Use NPU rather than NMU | arithmetic models · Simple arithmetic task (÷) | testing error: 11.399 ± 0.035 -> 0.23 ± 0.13 | Improved |
| Use NPU rather than NMU | arithmetic models · Simple arithmetic task (√) | testing error: 0.16 ± 0.002 -> 0.031 ± 0.025 | Improved |
| Use NMU rather than NPU | arithmetic models · Large scale arithmetic task (+) | testing MSE: 0.092 ± 0.031 -> 0.00602 ± 0.00019 | Improved |
| Use NMU rather than NPU | arithmetic models · Large scale arithmetic task (×) | testing MSE: 4.28 ± 0.9 -> 1.7 ± 1.4 | Improved |
| Use NPU rather than NMU | arithmetic models · Large scale arithmetic task (÷) | testing MSE: 1.622 ± 0.081 -> 1.0e-7 ± 1.0e-7 | Improved |
| Use NPU rather than NMU | arithmetic models · Large scale arithmetic task (√) | testing MSE: 10.96 ± 0.89 -> 0.054 ± 0.0078 | Improved |
| Add the relevance gate (ablation) | NaiveNPU · Large scale arithmetic task (+, ×, ÷, √) | testing MSE: NaiveNPU: 161.65 ± 0.11, 3750.0 ± 870.0, 5.4e17 ± 5.4e17, 9.3e8 ± 9.3e8; NPU: 0.092 ± 0.031, 4.28 ± 0.9, 1.0e-7 ± 1.0e-7 | Improved |
| Remove imaginary parameters from NPU (ablation) | NPU · Large scale arithmetic task (×) | testing MSE: 4.28 ± 0.9 -> 3.09 ± 0.74 | Improved |
| Remove imaginary parameters from NPU (ablation) | NPU · Large scale arithmetic task (÷) | testing MSE: 1.0e-7 ± 1.0e-7 -> 1.4e-6 ± 4.0e-7 | Hurt |
Closest papers by meaning
- Learning to Add, Multiply, and Execute Algorithmic Instructions Exactly with Neural Networks (NeurIPS 2025 · similarity 0.83)
- Emergence in non-neural models: grokking modular arithmetic via average gradient outer product (ICML 2025 · similarity 0.83)
- A Fast Convoluted Story: Scaling Probabilistic Inference for Integer Arithmetics (NeurIPS 2024 · similarity 0.82)
- Graph Neural Networks and Arithmetic Circuits (NeurIPS 2024 · similarity 0.82)
- Floating-Point Neural Networks Can Represent Almost All Floating-Point Functions (ICML 2025 · similarity 0.82)
- Arithmetic Without Algorithms: Language Models Solve Math with a Bag of Heuristics (ICLR 2025 · similarity 0.82)
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.
