Research map / Label-efficient and robust learning / Recognition and long-tail
Evaluating Capability of Deep Neural Networks for Image Classification via Information Plane
Hao Cheng, Dongze Lian, Shenghua Gao, Yanlin Geng
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 |
|---|---|---|---|
| increase depth from CNN-2 to CNN-16 (VGG) (ablation) | CNN on CIFAR-10 · CIFAR-10 validation | accuracy: 0.45 -> 0.89 | Improved |
| increase depth from CNN-9 to CNN-16 (ablation) | CNN on CIFAR-10 · CIFAR-10 validation | accuracy: 0.77 -> 0.89 | Improved |
| increase number of conv layers from CNN-2 to CNN-16 (ablation) | CNN trained by SGD · CIFAR-10 training set | inversion percentage: 0.56 -> 0.96 | Improved |
| train CNN-9 with SGD instead of BGD | CNN-9 · CIFAR-10 training set | inversion percentage: 0.821 -> 0.865 | Improved |
| train linear network with SGD instead of BGD | Linear network · CIFAR-10 training set | inversion percentage: 0.594 -> 0.755 | Improved |
| use CNN-9 instead of a linear network (SGD) | CNN-9 · CIFAR-10 training set | inversion percentage: 0.755 -> 0.865 | Improved |
| replace FC-3 with CNN-4 | FC-3 · MNIST validation (Table 4 models) | accuracy: 0.983 -> 0.994 | Improved |
| replace FC-3 with CNN-4 | FC-3 · CIFAR-10 validation (Table 4 models) | accuracy: 0.534 -> 0.705 | Improved |
| increase FC depth from 6 to 9 layers (ablation) | FC-6 · CIFAR-10 validation (Table 4 models) | accuracy: 0.575 -> 0.56 | Hurt |
| increase FC depth from 3 to 9 layers (ablation) | FC-3 · MNIST validation (Table 4 models) | accuracy: 0.983 -> 0.988 | Improved |
| increase number of samples from 100 to 600 for inversion percentage | VGG-16 · VGG-16 validation set | inversion percentage: 0.905 -> 0.924 | No clear effect |
Closest papers by meaning
- Learning Representations for Neural Network-Based Classification Using the Information Bottleneck Principle (TPAMI 2019 · similarity 0.87)
- Information Bottleneck and Aggregated Learning (TPAMI 2023 · similarity 0.85)
- Information Bottleneck Analysis of Deep Neural Networks via Lossy Compression (ICLR 2024 · similarity 0.85)
- Quantifying the Knowledge in a DNN to Explain Knowledge Distillation for Classification (TPAMI 2022 · similarity 0.85)
- DNNs as Layers of Cooperating Classifiers (AAAI 2020 · similarity 0.85)
- Quantification and Analysis of Layer-wise and Pixel-wise Information Discarding (ICML 2022 · 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.
