Research map / Architectures and efficiency / Quantization
Straightening Out the Straight-Through Estimator: Overcoming Optimization Challenges in Vector Quantized Networks
Minyoung Huh, Brian Cheung, Pulkit Agrawal, Phillip Isola
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 |
|---|---|---|---|
| Add the full combination of affine re-parameterization, synchronized updates, alternating optimization, and replacement | VQ AlexNet, ResNet18, and ViT(T) · ImageNet100 | Accuracy: 47.2 -> 57.9; 64.1 -> 71.0; 48.6 -> 56.7 | Improved |
| Add affine re-parameterization (ablation) | VQ AlexNet, ResNet18, and ViT(T) · ImageNet100 | Accuracy: 47.2 -> 55.7; 64.1 -> 70.4; 48.6 -> 52.8 | Improved |
| Add synchronized commitment updates (ablation) | VQ AlexNet, ResNet18, and ViT(T) · ImageNet100 | Accuracy: 47.2 -> 49.2; 64.1 -> 66.0; 48.6 -> 51.3 | Improved |
| Add alternating optimization (ablation) | VQ AlexNet, ResNet18, and ViT(T) · ImageNet100 | Accuracy: 47.2 -> 52.8; 64.1 -> 67.5; 48.6 -> 53.1 | Improved |
| Add l2 normalization | VQ classification models · ImageNet100 | Accuracy | Hurt |
| Add affine re-parameterization, alternating optimization, and replacement | VQVAE · CIFAR10 | MSE: 5.65 -> 1.81 | Improved |
| Also add l2 normalization to the proposed VQVAE combination (ablation) | VQVAE + Affine + OPT + Replace · CIFAR10 | MSE: 1.81 -> 1.74 | Improved |
| Add affine re-parameterization, alternating optimization, and replacement | VQVAE · CelebA | MSE: 10.02 -> 3.84 | Improved |
| Also add l2 normalization to the proposed VQVAE combination (ablation) | VQVAE + Affine + OPT + Replace · CelebA | MSE: 3.84 -> 4.42 | Hurt |
| Add affine re-parameterization and alternating optimization | MaskGIT · CelebA | FID: 90.4 -> 74.8 | Improved |
| Increase alternating optimization inner steps from 1 to 8 (ablation) | AlexNet · ImageNet100 | Accuracy: 52.54 (+5.58) -> 58.03 (+11.09) | Improved |
| Increase alternating optimization outer steps (ablation) | AlexNet · ImageNet100 | Accuracy: 55.13 (+8.19) -> 55.09 (+8.15); 57.45 (+10.51) -> 55.29 (+8.35) | No clear effect |
Closest papers by meaning
- And the Bit Goes Down: Revisiting the Quantization of Neural Networks (ICLR 2020 · similarity 0.88)
- Addressing Representation Collapse in Vector Quantized Models with One Linear Layer (ICCV 2025 · similarity 0.88)
- Learning to Quantize for Training Vector-Quantized Networks (ICML 2025 · similarity 0.87)
- Qinco2: Vector Compression and Search with Improved Implicit Neural Codebooks (ICLR 2025 · similarity 0.87)
- Restructuring Vector Quantization with the Rotation Trick (ICLR 2025 · similarity 0.86)
- Effective Interplay between Sparsity and Quantization: From Theory to Practice (ICLR 2025 · similarity 0.86)
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.
