Research map / Optimization / Optimizers
KOALA: A Kalman Optimization Algorithm with Loss Adaptivity
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 momentum dynamics with κ=0.90 instead of vanilla dynamics (ablation) | KOALA · CIFAR-100, ResNet-18 | Top-1 error: 24.17 -> 23.38 | Helped |
| Use momentum rate κ=0.50 instead of κ=0.90 (ablation) | KOALA-M · CIFAR-100, ResNet-18 | Top-1 error: 23.38 -> 27.20 | Hurt |
| Adapt state noise Qk with κ=0.90 (ablation) | KOALA-M · CIFAR-100, ResNet-18 | Top-1 error: 23.38 -> 23.39 | No clear effect |
| Adapt state noise Qk with κ=0.50 (ablation) | KOALA-M · CIFAR-100, ResNet-18 | Top-1 error: 27.20 -> 28.25 | Hurt |
| Vary posterior covariance initialization (ablation) | KOALA-V and KOALA-M · CIFAR-100, ResNet-18 | Top-1 error: KOALA-V: 24.42, 24.17, 24.69; σ²x,0: 23.67, 23.39, 23.82; σ²p,0: 23.37, 23.39, 24.24 | No clear effect |
| Vary the batch size from 32 to 256 (ablation) | KOALA-M · CIFAR-100, ResNet-18 | Top-1 error: 24.59, 23.11, 23.39, 24.34 | No clear effect |
| Use Xavier-Normal initialization (ablation) | KOALA-M · CIFAR-100, ResNet-18 | Top-1 error: SGD 26.71; KOALA-M 23.34 | Helped |
| Use Xavier-Uniform initialization (ablation) | KOALA-M · CIFAR-100, ResNet-18 | Top-1 error: SGD 26.90; KOALA-M 23.40 | Helped |
| Use Kaiming-Uniform initialization (ablation) | KOALA-M · CIFAR-100, ResNet-18 | Top-1 error: SGD 27.82; KOALA-M 23.35 | Helped |
| Use Orthogonal initialization (ablation) | KOALA-M · CIFAR-100, ResNet-18 | Top-1 error: SGD 26.83; KOALA-M 23.27 | Helped |
| Train with KOALA-M instead of SGD | ResNet-18 · CIFAR-100, 100 epochs | Top-1 error: 23.50 -> 23.38 | Helped |
| Train with KOALA-M instead of SGD | ResNet-18 · CIFAR-100, 200 epochs | Top-1 error: 22.44 -> 22.22 | Helped |
Closest papers by meaning
- KOALA++: Efficient Kalman-Based Optimization with Gradient-Covariance Products (NeurIPS 2025 · similarity 0.93)
- Optimization or Architecture: How to Hack Kalman Filtering (NeurIPS 2023 · similarity 0.87)
- Implicit Maximum a Posteriori Filtering via Adaptive Optimization (ICLR 2024 · similarity 0.86)
- Accelerating Neural Network Training with Augmented Koopman Dynamics (NeurIPS 2026 · similarity 0.86)
- Predictively-Oriented Kalman Filtering (NeurIPS 2026 · similarity 0.85)
- Deep-Koopman-KANDy: Dictionary Discovery for Deep-Koopman Operators with Kolmogorov-Arnold Networks for Dynamics (NeurIPS 2026 · 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.