Research map / Optimization / Optimizers
Approximate Fisher Information Matrix to Characterize the Training of Deep Neural Networks
Zhibin Liao, Tom W. Drummond, Ian D. Reid, Gustavo Carneiro
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 |
|---|---|---|---|
| increase batch size over training (s32-to-512) instead of decreasing (s512-to-32) (ablation) | ResNet dynamic sampling · CIFAR-10 | test error (%): 5.56 -> 5.07 | Improved |
| dynamic sampling instead of best fixed batch size | ResNet s32 · CIFAR-10 | test error (%): 4.78 -> 4.90 | No clear effect |
| dynamic sampling instead of best fixed batch size | ResNet s32 · CIFAR-10 | training time (h): 7.7 -> 7.0 | Improved |
| multi-step dynamic sampling (-MS) instead of whole-training dynamic sampling (ablation) | ResNet · CIFAR-10 | test error (%): 4.90 -> 4.76 | Improved |
| multi-step dynamic sampling (-MS) instead of whole-training dynamic sampling (ablation) | DenseNet · CIFAR-10 | test error (%): 5.03 -> 4.63 | Improved |
| multi-step dynamic sampling (-MS) instead of whole-training dynamic sampling (ablation) | DenseNet · CIFAR-100 | test error (%): 23.68 -> 23.92 | Hurt |
| multi-step dynamic sampling (-MS) instead of whole-training dynamic sampling (ablation) | ResNet · MNIST | test error (%): 0.39 -> 0.34 | Improved |
| multi-step dynamic sampling (-MS) instead of fixed s32 | DenseNet · CIFAR-10 | test error (%): 4.96 -> 4.63 | Improved |
| dynamic sampling instead of fixed batch | DenseNet s128 · SVHN | test error (%): 1.93 -> 2.00 | No clear effect |
| dynamic sampling instead of fixed batch | DenseNet · MNIST | training time (h): 0.26 -> 0.06 | Improved |
| dynamic sampling instead of fixed batch | DenseNet · CIFAR-10 | training time (h): 16.1 -> 14.1 | Improved |
| use small learning rate 0.05 instead of 0.4 with batch size 8 (ablation) | ResNet s8 · CIFAR-10 | test error (%): 17.23 -> 5.09 | Improved |
Closest papers by meaning
- Learning Rates as a Function of Batch Size: A Random Matrix Theory Approach to Neural Network Training (JMLR 2022 · similarity 0.88)
- An Efficient Fisher Matrix Approximation Method for Large-Scale Neural Network Optimization (TPAMI 2022 · similarity 0.87)
- An Improved Empirical Fisher Approximation for Natural Gradient Descent (NeurIPS 2024 · similarity 0.85)
- Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights (CVPR 2025 · similarity 0.85)
- Fishers for Free? Approximating the Fisher Information Matrix by Recycling the Squared Gradient Accumulator (ICML 2025 · similarity 0.85)
- A Layer-Wise Natural Gradient Optimizer for Training Deep Neural Networks (NeurIPS 2024 · similarity 0.84)
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.
