Research map / Architectures and efficiency / CNN design
Causes and Effects of Unanticipated Numerical Deviations in Neural Network Inference Frameworks
Alexander Schlögl, Nora Hofer, Rainer Böhme
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 |
|---|---|---|---|
| cast model from float32 to float16 (ablation) | Cifar10-R18 · 75 platforms | number of EQCs: 9 -> 3 | Improved |
| cast model from float32 to float64 (ablation) | Cifar10-R18 · 75 platforms | number of EQCs: 9 -> 4 | Improved |
| increase CPU core count from 2 to 4 (ablation) | Intel Ice Lake · Cifar10-R18 | number of EQCs: 5 -> 4 | Improved |
| increase CPU core count from 8 to 16 (ablation) | Intel platforms · Cifar10-R18 | number of EQCs: 4 -> 3 | Improved |
| increase number of core counts beyond 16 (ablation) | Intel platforms · Cifar10-R18 | number of EQCs | No clear effect |
| increase number of multiplications per convolution layer (ablation) | convolution layer · remaining precision | remaining precision: 23 -> 0 | Hurt |
| increase number of multiplications from 10^6-10^7 to 10^8 (ablation) | convolution layer · number of EQCs | number of EQCs | Mixed |
| change weight/input distribution (original, Glorot, Gaussian; captured, permuted, uniform inputs) (ablation) | conv2d_11 · dual-core GCP instances | number of EQCs | No clear effect |
| enable XLA compute-graph optimization (not activated) | TensorFlow · CPU/GPU inference | number of EQCs | No clear effect |
| different device placement | TensorFlow · CPU/GPU inference | number of EQCs | No clear effect |
| AMD Rome/Milan or Sandy/Ivy Bridge at larger core counts | CPU inference · Cifar10-R18 | number of EQCs | No clear effect |
| different GPU model (GTX 1650 vs RTX 2070) | Cifar10-R18 · GPU inference | number of EQCs: 1 -> 2 | Hurt |
Closest papers by meaning
- Surprising Instabilities in Training Deep Networks and a Theoretical Analysis (NeurIPS 2022 · similarity 0.86)
- Understanding and Mitigating Numerical Sources of Nondeterminism in LLM Inference (NeurIPS 2025 · similarity 0.85)
- A PDE-based Explanation of Extreme Numerical Sensitivities and Edge of Stability in Training Neural Networks (JMLR 2024 · similarity 0.85)
- Initial Guessing Bias: How Untrained Networks Favor Some Classes (ICML 2024 · similarity 0.84)
- Effective Interplay between Sparsity and Quantization: From Theory to Practice (ICLR 2025 · similarity 0.83)
- Don’t Think It Twice: Exploit Shift Invariance for Efficient Online Streaming Inference of CNNs (AAAI 2025 · 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.
