Research map / Architectures and efficiency / RNNs and normalization
Lookaround Optimizer: $k$ steps around, 1 step average
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 |
|---|---|---|---|
| Train augmentation-specific models from a shared point and average their weights iteratively | SGD training · CIFAR100 | Top1: 80.06 -> 81.60 | Helped |
| Use iterative weight averaging with augmentation-specific models | ResNet50 on ImageNet · ImageNet validation | Top-1: 75.97 -> 77.32 | Helped |
| Add SAM to Lookaround | ResNet50 pretrained · CIFAR100 | accuracy: 85.20 -> 86.21 | Helped |
| Add SAM to Lookaround | ViT-B/16 pretrained · CIFAR100 | accuracy: 92.21 -> 92.54 | Helped |
| Add SWA to Lookaround | ResNet18 · Stanford Cars | Top-1 Acc: 90.76 -> 91.04 | Helped |
| Use data augmentation together with weight averaging (ablation) | Lookaround, ResNet50 · CIFAR10 | Acc (%): 95.84 -> 96.59 | Helped |
| Use data augmentation together with weight averaging (ablation) | Lookaround, ResNet50 · CIFAR100 | Acc (%): 78.81 -> 81.60 | Helped |
| Add weight averaging to data augmentation (ablation) | Lookaround, ResNet50 · CIFAR100 | Acc (%): 79.61 -> 81.60 | Helped |
| Add data augmentation to weight averaging (ablation) | Lookaround, ResNet50 · CIFAR100 | Acc (%): 79.49 -> 81.60 | Helped |
| Increase the number of data augmentations from 1 to 2 (ablation) | Lookaround, ResNet50 · CIFAR100 | Top-1 (%): 78.20 -> 80.82 | Helped |
| Increase the number of data augmentations from 2 to 3 (ablation) | Lookaround, ResNet50 · CIFAR100 | Top-1 (%): 80.82 -> 81.60 | Helped |
| Increase the number of data augmentations from 3 to 4 (ablation) | Lookaround, ResNet50 · CIFAR100 | Top-1 (%): 81.60 -> 81.19 | Hurt |
Closest papers by meaning
- When, Where and Why to Average Weights? (ICML 2025 · similarity 0.88)
- Diverse Weight Averaging for Out-of-Distribution Generalization (NeurIPS 2022 · similarity 0.88)
- Trainable Weight Averaging: Efficient Training by Optimizing Historical Solutions (ICLR 2023 · similarity 0.87)
- Why Do We Need Weight Decay in Modern Deep Learning? (NeurIPS 2024 · similarity 0.85)
- Gradient-Weight Alignment as a Train-Time Proxy for Generalization in Classification Tasks (NeurIPS 2025 · similarity 0.84)
- Rethinking Weight Decay for Robust Fine-Tuning of Foundation Models (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.