Unsupervised Deep Learning by Neighbourhood Discovery
Authors: Jiabo Huang, Qi Dong, Shaogang Gong, Xiatian Zhu
ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on image classification show the performance advantages of the proposed method over the state-of-the-art unsupervised learning models on six benchmarks including both coarse-grained and fine-grained object image categorisation. Extensive experiments are conducted on four coarse-grained (CIFAR10 and CIFAR100 (Krizhevsky & Hinton, 2009), SVHN (Netzer et al., 2011), Image Net (Russakovsky et al., 2015)) and two fine-grained (CUB200-2011 (Wah et al., 2011) and Stanford Dogs (Khosla et al., 2011)) object image classification datasets. The results show the advantages of our AND method over a wide variety of existing state-ofthe-art unsupervised deep learning models. |
| Researcher Affiliation | Collaboration | Jiabo Huang 1 Qi Dong 1 Shaogang Gong 1 Xiatian Zhu 2 1Queen Mary University of London 2Vision Semantics Limited. |
| Pseudocode | Yes | Algorithm 1 Neighbourhood discovery. |
| Open Source Code | Yes | Code is available at https://github.com/raymond-sci/AND. |
| Open Datasets | Yes | Datasets. We used 6 image classification benchmarks for evaluating our model (Fig 3). CIFAR10(/100) (Krizhevsky & Hinton, 2009): An image dataset with 50,000/10,000 train/test images from 10 (/100) object classes. SVHN (Netzer et al., 2011): A Street View House Numbers dataset including 10 classes of digit images. Image Net (Russakovsky et al., 2015): A large 1,000 classes object dataset with 1.2 million images for training and 50,000 for test. CUB2002011 (Wah et al., 2011): A fine-grained dataset containing 5,994/5,794 train/test images of 200 bird species. Stanford Dogs (Khosla et al., 2011): A fine-grained dataset with 12,000/8,580 train/test images of 120 dog breeds. |
| Dataset Splits | No | The paper specifies train and test splits for datasets (e.g., '50,000/10,000 train/test images' for CIFAR10/100), but does not explicitly mention validation splits or how they are derived for any of the datasets used. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU models, CPU models, memory sizes, or cloud instance types) used for running the experiments. |
| Software Dependencies | No | The paper mentions using 'Alex Net' as the neural network architecture, but it does not specify any software dependencies (e.g., programming languages, libraries, frameworks) with version numbers that would be needed to replicate the experiments. |
| Experiment Setup | Yes | To train AND models, we set the learning rate to 0.03 which was further scaled down by 0.1 every 40 epochs after the first 80 epochs. We used the batch size of 256 for Image Net and 128 for others. We set the epoch to 200 per round. We fixed the feature length to 128. We applied the SGD with Nesterov momentum at 0.9. Our model usually converges with R=4 rounds. We set η=0.5 in Eq (8) for feature update. We set k = 1 (Eq (1)) for exploring the most local neighbourhoods. |