Research map / Optimization / Optimizers
Enhancing Optimization Robustness in 1-bit Neural Networks through Stochastic Sign Descent
NianHui Guo, Hong Guo, Christoph Meinel, Haojin Yang
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 |
|---|---|---|---|
| replace Adam with the Diode sign-descent optimizer on BNext-18† | BNext-18† · ImageNet ILSVRC2012 val | Top-1 accuracy (%): 67.5 -> 68.2 | Improved |
| replace Adam with Diode on ReActNet-A without two-stage training | ReActNet-A · ImageNet ILSVRC2012 val | Top-1 accuracy (%): 69.4 -> 69.6 | Improved |
| replace Adam with Diode on binary BERT (BiT‡) without data augmentation | BiT · GLUE | average score: 71.0 -> 74.7 | Improved |
| use BOP or Gradient Filter latent-weight-free optimizers instead of Adam on binary BERT (ablation) | BiT · GLUE | average score: 71.0 -> 61.9 (BOP), 62.5 (Gradient Filter) | Hurt |
| initialize Diode from pre-trained BiT weights (ablation) | BiT · GLUE | average score: 74.7 -> 75.7 | Improved |
| add data augmentation for Diode on BiT† (ablation) | BiT · GLUE | average score: 75.7 -> 78.8 | Improved |
| increase long-term EMA ratio beta from 0.999 to 0.9999 (ablation) | Diode · CIFAR100 BNext-18† | Top-1 accuracy (%): 72.01 -> 74.51 (alpha=0.99) | Improved |
| vary initial learning rate eta from 1.0 to 1e-4 (ablation) | Diode · CIFAR100 BNext-18† | Top-1 accuracy (%): 74.50 vs 74.45 | No clear effect |
| use CosineLR eta decay instead of step or exponential decay | Diode · CIFAR100 BNext-18† | validation error | Improved |
| train with batch size 4096 using Diode instead of best Adam (B=512) (ablation) | BNext-18† · ImageNet ILSVRC2012 val | Top-1 accuracy (%): +0.96% with 8x fewer iterations | Improved |
| use Diode with small batch size (below 128) | BNext-18† · ImageNet ILSVRC2012 val | Top-1 accuracy (%) | Hurt |
Closest papers by meaning
- Sub-Bit Neural Networks: Learning To Compress and Accelerate Binary Neural Networks (ICCV 2021 · similarity 0.88)
- OvSW: Overcoming Silent Weights for Accurate Binary Neural Networks (ECCV 2024 · similarity 0.87)
- BD-Net: Has Depth-Wise Convolution Ever Been Applied in Binary Neural Networks? (AAAI 2026 · similarity 0.87)
- Resilient Binary Neural Network (AAAI 2023 · similarity 0.86)
- An Empirical study of Binary Neural Networks' Optimisation (ICLR 2019 · similarity 0.86)
- Training Binary Neural Networks via Gaussian Variational Inference and Low-Rank Semidefinite Programming (NeurIPS 2024 · similarity 0.85)
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.
