Research map / Label-efficient and robust learning / Recognition and long-tail
Discriminative Unsupervised Feature Learning with Exemplar Convolutional Neural Networks
Alexey Dosovitskiy, Philipp Fischer, Jost Tobias Springenberg, Martin Riedmiller, Thomas Brox
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 surrogate classes up to about 8000 (ablation) | Exemplar-CNN 64c5-64c5-128f · STL-10, CIFAR-10, Caltech-101 | classification accuracy (%): 50 -> 8000 classes helps; beyond 8000 no change or decrease | Mixed |
| add more samples per surrogate class (1 to 300) (ablation) | Exemplar-CNN 64c5-64c5-128f · STL-10 | classification accuracy (%): saturates at around 100 samples | Improved |
| add translation, color and contrast transformations to seed augmentation (ablation) | Exemplar-CNN 64c5-64c5-128f · STL-10, CIFAR-10, Caltech-101 | classification accuracy (%) | Improved |
| add rotation and scaling transformations (ablation) | Exemplar-CNN 64c5-64c5-128f · STL-10, CIFAR-10, Caltech-101 | classification accuracy (%) | No clear effect |
| add occlusion, affine transformation, Gaussian noise to augmentation (ablation) | Exemplar-CNN · classification | classification accuracy (%) | No clear effect |
| add clustering to build surrogate classes (ablation) | Exemplar-CNN 92c5-256c5-512c5-1024f · STL-10 | classification accuracy (%): 74.2 -> 75.4 | Improved |
| train seed patches from the test dataset itself instead of STL-10 (Caltech-101 seeds, Caltech-101 test) (ablation) | Exemplar-CNN 64c5-64c5-128f · Caltech-101 | classification accuracy (%): 79.8 -> 80.0 | No clear effect |
| train on CIFAR-10 seed patches instead of STL-10 (ablation) | Exemplar-CNN 64c5-64c5-128f · STL-10 | classification accuracy (%): 67.1 -> 64.5 | Hurt |
| add a fourth convolutional layer (ablation) | Exemplar-CNN 64c5-64c5-64c5-128f · STL-10 | classification accuracy (%): 68.5 -> 64.7 | Hurt |
| add a third convolutional layer (ablation) | Exemplar-CNN 64c5-64c5-128f · STL-10 | classification accuracy (%): 67.1 -> 68.5 | Improved |
| widen layers from 64-64-128 to 128-256-512 (ablation) | Exemplar-CNN 64c5-64c5-128f · STL-10 | classification accuracy (%): 67.1 -> 71.2 | Improved |
| use 7x7 instead of 5x5 first-layer filters (ablation) | Exemplar-CNN 64c5-64c5-128f · STL-10 | classification accuracy (%): 67.1 -> 66.3 | No clear effect |
Closest papers by meaning
- Unsupervised Deep Learning via Affinity Diffusion (AAAI 2020 · similarity 0.91)
- Unsupervised Deep Learning by Neighbourhood Discovery (ICML 2019 · similarity 0.90)
- Unsupervised Representation Learning by Predicting Image Rotations (ICLR 2018 · similarity 0.90)
- Unsupervised Learning of Discriminative Attributes and Visual Representations (CVPR 2016 · similarity 0.89)
- Convolutional Patch Representations for Image Retrieval: An Unsupervised Approach (IJCV 2016 · similarity 0.87)
- PICNN: A Pathway towards Interpretable Convolutional Neural Networks (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.
