Research map / Optimization / Distributed and compressed
Non-reversible Parallel Tempering for Deep Posterior Approximation
Wei Deng, Qian Zhang, Qi Feng, Faming Liang, Guang Lin
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 DEO★ windowing instead of vanilla DEO window size 1 (ablation) | mSGD parallel tempering · CIFAR100, ResNet20 | NLL: 7964±23 -> 7741±67 | Improved |
| Use DEO★ windowing instead of vanilla DEO window size 1 (ablation) | mSGD parallel tempering · CIFAR100, ResNet32 | NLL: 7152±41 -> 7019±35 | Improved |
| Use DEO★ windowing instead of vanilla DEO window size 1 (ablation) | mSGD parallel tempering · CIFAR100, ResNet56 | NLL: 6534±26 -> 6439±32 | Improved |
| Use DEO★ windowing instead of vanilla DEO window size 1 (ablation) | mSGD parallel tempering · CIFAR100, ResNet20 | ACC (%): 76.84±0.12 -> 77.37±0.16 | Improved |
| Use DEO★ windowing instead of vanilla DEO window size 1 (ablation) | mSGD parallel tempering · CIFAR100, ResNet32 | ACC (%): 79.34±0.15 -> 79.54±0.12 | Improved |
| Use DEO★ windowing instead of vanilla DEO window size 1 (ablation) | mSGD parallel tempering · CIFAR100, ResNet56 | ACC (%): 81.72±0.12 -> 82.02±0.15 | Improved |
| Increase window size from 1 to 8 (ablation) | DEO parallel tempering · Multimodal simulation, 16 chains | round trips per 1,000 iterations: 18 -> 45 | Improved |
| Use target swap rate 0.2 instead of 0.4 (ablation) | DEO★-SGD×P16 · Multimodal simulation | posterior approximation | Hurt |
| Use target swap rate 0.6 instead of 0.4 (ablation) | DEO★-SGD×P16 · Multimodal simulation | posterior approximation | Hurt |
| Use DEO★ instead of vanilla DEO swaps (ablation) | SGD-based parallel tempering · Multimodal simulation, 16 chains | posterior approximation | Improved |
| Replace single-chain cyclical SGLD with DEO★-SGD parallel tempering | posterior approximation · Multimodal simulation | posterior approximation | Improved |
| Use DEO★ windowing instead of DEO window size 1 (ablation) | parallel tempering · CIFAR100, ResNet20 | round trips per 100 epochs: almost 2 times | Improved |
Closest papers by meaning
- A connection between Tempering and Entropic Mirror Descent (ICML 2024 · similarity 0.87)
- Constrained Exploration via Reflected Replica Exchange Stochastic Gradient Langevin Dynamics (ICML 2024 · similarity 0.85)
- Communication-Efficient Diffusion Denoising Parallelization via Reuse-then-Predict Mechanism (NeurIPS 2025 · similarity 0.85)
- Provable Posterior Sampling with Denoising Oracles via Tilted Transport (NeurIPS 2024 · similarity 0.85)
- Progressive Tempering Sampler with Diffusion (ICML 2025 · similarity 0.84)
- Provable Convergence and Limitations of Geometric Tempering for Langevin Dynamics (ICLR 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.
