Research map / Label-efficient and robust learning / Recognition and long-tail
Joint Unsupervised Learning of Deep Representations and Image Clusters
Check 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 |
|---|---|---|---|
| Jointly learn CNN representations and clusters, then re-cluster with the final representation rather than using the recurrent process output directly (ablation) | JULE · Image clustering across COIL20, COIL100, USPS, MNIST-test, MNIST-full, UMist, FRGC, CMU-PIE, and YTF | NMI: OURS-RC is better on COIL100, USPS, MNIST-test, MNIST-full, FRGC, and CMU-PIE; tied on COIL20 and YTF; lower on UMist | Mixed |
| Use the joint-learning method instead of the strongest listed raw-intensity baseline | JULE · Image clustering across the nine datasets | NMI: 0.954/0.911 -> 1.000 (COIL20); 0.964 -> 0.985 (COIL100); 0.854/0.824 -> 0.913 (USPS); 0.864/0.844 -> 0.915 (MNIST-test); | Helped |
| Transfer a learned representation instead of using raw input pixels | JULE representations · COIL20 → COIL100 | NMI: 0.924 -> 0.939 | Helped |
| Transfer a learned representation instead of using raw input pixels | JULE representations · COIL100 → COIL20 | NMI: 0.944 -> 0.957 | Helped |
| Transfer a learned representation instead of using raw input pixels | JULE representations · MNIST-test → USPS | NMI: 0.874 -> 0.908 | Helped |
| Transfer a learned representation instead of using raw input pixels | JULE representations · USPS → MNIST-test | NMI: 0.872 -> 0.886 | Helped |
| Combine conv1 and conv2 features rather than using either layer alone (ablation) | CIFAR-10 codebook pipeline · CIFAR-10 test, 50k training samples | accuracy: 74.68% -> 78.55% | Helped |
| Use the conv1&2 representation instead of K-means features | CIFAR-10 codebook pipeline · CIFAR-10 test, 50k training samples | accuracy: 76.59% -> 78.55% | Helped |
| Fine-tune the unsupervised model using predicted cluster labels rather than ground-truth supervised labels | Face verification model · LFW restricted, training on 30k YTF samples | accuracy: 0.748 -> 0.750 | Helped |
| Fine-tune the unsupervised model using predicted cluster labels rather than ground-truth supervised labels | Face verification model · LFW restricted, training on 10k, 20k, 50k, and 100k YTF samples | accuracy: 10k: 0.737 -> 0.728; 20k: 0.746 -> 0.743; 50k: 0.764 -> 0.762; 100k: 0.770 -> 0.767 | Mixed |
Closest papers by meaning
- Online Deep Clustering for Unsupervised Representation Learning (CVPR 2020 · similarity 0.90)
- ClusterFit: Improving Generalization of Visual Representations (CVPR 2020 · similarity 0.89)
- Clustering-friendly Representation Learning via Instance Discrimination and Feature Decorrelation (ICLR 2021 · similarity 0.89)
- Unsupervised Deep Embedding for Clustering Analysis (ICML 2016 · similarity 0.89)
- Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels (ICLR 2024 · similarity 0.87)
- Image Clustering via the Principle of Rate Reduction in the Age of Pretrained Models (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.