Research map / Optimization / Optimizers
Memory Efficient Adaptive Optimization
Rohan Anil, Vineet Gupta, Tomer Koren, Yoram Singer
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 SM3 instead of Adam | Transformer-Big · WMT'14 en→fr, batch 384 | BLEU: 38.96 ± 0.002 -> 39.81 ± 0.002 | Improved |
| Use SM3 instead of Adafactor | Transformer-Big · WMT'14 en→fr, batch 768 | BLEU: 39.65 ± 0.002 -> 40.50 ± 0.001 | Improved |
| Increase batch size from 384 to 768 using SM3 | SM3 on Transformer-Big · WMT'14 en→fr | BLEU: 39.81 ± 0.002 -> 40.50 ± 0.001 | Improved |
| Use SM3 instead of Adam | BERT-Large · Wikipedia+BooksCorpus | Masked-LM test accuracy | No clear effect |
| Increase batch size from 1024 to 2048 using SM3 | SM3 on BERT-Large · Wikipedia+BooksCorpus | Masked-LM test accuracy | Improved |
| Use a double batch size with SM3 | BERT-Large · Wikipedia+BooksCorpus, target 70% Masked-LM accuracy | wall-time to target: almost 35% less wall-time | Improved |
| Use SM3 instead of SGD with momentum | AmoebaNet-D · ImageNet | top-1 accuracy: 78.71% | Improved |
| Use SM3-II rather than SM3-I (ablation) | SM3 · Adagrad accumulator approximation | approximation tightness | Improved |
Closest papers by meaning
- Sophia: A Scalable Stochastic Second-order Optimizer for Language Model Pre-training (ICLR 2024 · similarity 0.90)
- Unleashing the Potential of Large Language Models as Prompt Optimizers: Analogical Analysis with Gradient-based Model Optimizers (AAAI 2025 · similarity 0.89)
- LDAdam: Adaptive Optimization from Low-Dimensional Gradient Statistics (ICLR 2025 · similarity 0.89)
- AdaRankGrad: Adaptive Gradient Rank and Moments for Memory-Efficient LLMs Training and Fine-Tuning (ICLR 2025 · similarity 0.89)
- Small Batch Size Training for Language Models: When Vanilla SGD Works, and Why Gradient Accumulation is Wasteful (NeurIPS 2025 · similarity 0.89)
- Optimization Inspired Few-Shot Adaptation for Large Language Models (NeurIPS 2025 · similarity 0.88)
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.
