Research map / Detection and segmentation / Open-set and few-shot detection
Minimal Images in Deep Neural Networks: Fragile Object Recognition in Natural Images
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 |
|---|---|---|---|
| augment training with crops at FRI sizes | CIFAR-10 AlexNet · CIFAR-10 | FRI occurrence: more than 12% decrease for 12-pixel regions | Helped |
| add weight decay, dropout, and image distortions | CIFAR-10 AlexNet · CIFAR-10 | FRI occurrence: more than 12% decrease for 12-pixel regions | Helped |
| use dropout as the regularizer (ablation) | CIFAR-10 AlexNet · CIFAR-10 | FRI occurrence | Helped |
| increase the second-layer pooling region size (ablation) | CIFAR-10 AlexNet · embedded 12-pixel images | one-pixel shifts affecting output: 75% -> 10% | Helped |
| increase the second-layer pooling region size (ablation) | CIFAR-10 AlexNet · CIFAR-10, 12-pixel FRI regions | FRI occurrence: approximately 10% less; about 65% of the regions are FRIs | Helped |
| increase the second-layer pooling region size (ablation) | CIFAR-10 AlexNet · CIFAR-10 crops | accuracy | No clear effect |
| evaluate DNN FRI pairs instead of human minimal-image pairs | human observers · ImageNet DNN FRI pairs | recognition success-rate gap between neighboring crops: 14.5% for humans vs. 56% DNN confidence gap | No clear effect |
| use a multi-scale eccentricity-dependent architecture (ablation) | CIFAR-10 DNN · CIFAR-10 | FRI occurrence | Helped |
Closest papers by meaning
- Harmonizing the object recognition strategies of deep neural networks with humans (NeurIPS 2022 · similarity 0.90)
- Generalisation in humans and deep neural networks (NeurIPS 2018 · similarity 0.89)
- Performance-optimized deep neural networks are evolving into worse models of inferotemporal visual cortex (NeurIPS 2023 · similarity 0.88)
- Progress and Limitations of Deep Networks to Recognize Objects in Unusual Poses (AAAI 2023 · similarity 0.88)
- Deep Residual Learning for Image Recognition (CVPR 2016 · similarity 0.87)
- Harnessing Edge Information for Improved Robustness in Vision Transformers (AAAI 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.