Research map / Optimization / Optimizers
Learning compositional functions via multiplicative weight updates
Jeremy Bernstein, Jiawei Zhao, Markus Meister, Ming-Yu Liu, Anima Anandkumar, Yisong Yue
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 Madam instead of tuned Adam | ResNet-18 · CIFAR-10 | top-1 test error: 6.6\pm0.2 -> 7.8\pm0.2 | Hurt |
| Use Madam instead of tuned SGD | ResNet-18 · CIFAR-100 | top-1 test error: 29.1\pm0.2 -> 30.2\pm0.1 | Hurt |
| Use Madam instead of tuned SGD | ResNet-50 · ImageNet | top-1 test error: 24.1\pm0.1 -> 28.9\pm0.1 | Hurt |
| Use Madam instead of tuned Adam | BigGAN cGAN · CIFAR-10 | FID: 23.9\pm0.9 -> 19 \pm 2 | Improved |
| Use Madam instead of tuned SGD | Transformer · WikiText-2 | perplexity: 169.6\pm0.6 -> 173.3\pm0.6 | Hurt |
| Use 12-bit rather than FP32 Madam | Madam with ResNet-18 · CIFAR-10 | top-1 test error: 7.8\pm0.2 -> 7.0\pm0.1 | Improved |
| Use 12-bit rather than FP32 Madam | Madam with ResNet-18 · CIFAR-100 | top-1 test error: 30.2\pm0.1 -> 27.6\pm0.3 | Improved |
| Use 12-bit rather than FP32 Madam | Madam with ResNet-50 · ImageNet | top-1 test error: 28.9\pm0.1 -> 31.1\pm0.1 | Hurt |
| Use 12-bit rather than FP32 Madam | Madam with BigGAN cGAN · CIFAR-10 | FID: 19.3\pm0.7 -> 19.8\pm0.8 | No clear effect |
| Use 12-bit rather than FP32 Madam | Madam with Transformer · WikiText-2 | perplexity: 173.3\pm0.6 -> 182.3\pm0.6 | Hurt |
| Reduce weight precision to 8 bits | Madam · ImageNet | top-1 test error: 28.9\pm0.1 -> 50.5\pm0.5 | Hurt |
| Reduce weight precision to 8 bits | Madam · WikiText-2 | perplexity: 173.3\pm0.6 -> 262\pm2 | Hurt |
Closest papers by meaning
- Position: A Theory of Deep Learning Must Include Compositional Sparsity (ICML 2025 · similarity 0.87)
- A Second-Order Perspective on Model Compositionality and Incremental Learning (ICLR 2025 · similarity 0.86)
- How DNNs break the Curse of Dimensionality: Compositionality and Symmetry Learning (ICLR 2025 · similarity 0.85)
- CTSketch: Compositional Tensor Sketching for Scalable Neurosymbolic Learning (NeurIPS 2025 · similarity 0.84)
- Feature emergence via margin maximization: case studies in algebraic tasks (ICLR 2024 · similarity 0.84)
- Improving Deep Learning Speed and Performance Through Synaptic Neural Balance (AAAI 2025 · 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.
