Research map / Label-efficient and robust learning / Recognition and long-tail
Understanding Invariance via Feedforward Inversion of Discriminatively Trained Classifiers
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 the feedforward conditional GAN inversion method instead of D&B inversion | Robust ResNet-152 logits · ImageNet test reconstruction pairs | human preference: D&B: 12.5%; Ours: 87.5% | Helped |
| Use the feedforward conditional GAN inversion method instead of D&B inversion | Robust ResNet-152 logits · ImageNet test | LPIPS: 0.3092 -> 0.3138 | Hurt |
| Use robust rather than standard training | ResNet-152 logits and their reconstructions · ImageNet test reconstruction pairs | human preference: Non-robust: 23.5%; Robust: 76.5% | Helped |
| Use robust rather than standard training | ResNet-152 logits and their reconstructions · ImageNet test | LPIPS: 0.3755 -> 0.3138 | Helped |
| Use ResNet-152 rather than Inception-V3 | Standard classifier logits and their reconstructions · ImageNet test reconstruction pairs | human preference: ResNet: 52.7%; Inception: 47.3% | Mixed |
| Use ResNet-152 rather than Inception-V3 | Standard classifier logits and their reconstructions · ImageNet test | LPIPS: 0.3881 -> 0.3755 | Hurt |
| Use robust rather than standard training | ResNet-152 inversion · ImageNet, correctly classified samples | reconstruction classified identically to ground truth: 49.3% -> 54% | Helped |
| Use robust rather than standard training | ResNet-152 inversion · ImageNet, incorrectly classified samples | reconstruction classified identically to ground truth: 29.9% -> 35% | Helped |
| Use robust rather than standard training | ImageNet-trained ResNet-152 inversion · CIFAR-100, MNIST, and FashionMNIST | reconstruction fidelity | Helped |
| Resample generator noise while keeping logits fixed | Robust and non-robust ResNet-152 inversion · ImageNet | variation in reconstruction content | Mixed |
| Add Gaussian noise to logits | Robust rather than non-robust ResNet-152 inversion · ImageNet | reconstruction content stability: 42.33 vs. 43.78 | Helped |
| Invert correctly classified rather than incorrectly classified inputs | Robust ResNet-152 · ImageNet | reconstruction classified identically to ground truth: 35% -> 54% | Helped |
Closest papers by meaning
- Learning Invariances in Neural Networks from Training Data (NeurIPS 2020 · similarity 0.89)
- Do deep networks transfer invariances across classes? (ICLR 2022 · similarity 0.89)
- Grounding inductive biases in natural images: invariance stems from variations in data (NeurIPS 2021 · similarity 0.88)
- Tilt your Head: Activating the Hidden Spatial-Invariance of Classifiers (ICML 2024 · similarity 0.88)
- Unsupervised Adversarial Invariance (NeurIPS 2018 · similarity 0.88)
- From Bricks to Bridges: Product of Invariances to Enhance Latent Space Communication (ICLR 2024 · 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.