Unsupervised Deep Learning via Affinity Diffusion
Authors: Jiabo Huang, Qi Dong, Shaogang Gong, Xiatian Zhu11029-11036
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on object image classification and clustering show the performance superiority of the proposed method over the stateof-the-art unsupervised learning models using six common image recognition benchmarks including MNIST, SVHN, STL10, CIFAR10, CIFAR100 and Image Net. |
| Researcher Affiliation | Collaboration | 1Queen Mary University of London, 2Vision Semantics Limited |
| Pseudocode | Yes | Algorithm 1 Unsupervised deep learning via progressive affinity diffusion. Input: Training data I, training epochs Nep, iterations per epoch Nit; Output: A class discriminative CNN feature representation model; for epoch = 1 to Nep do Construct the k NN based affinity graph (Eq (1)); Search strongly connected subgraphs (Fig 3); for iter = 1 to Nit do Mini-batch feed-forward through the network; Objective loss computation (Eq (5)); Network back-propagation and model weights update; Memory feature refreshing (Eq (6)). end for end for |
| Open Source Code | No | The paper does not include any explicit statement about releasing the source code or a link to a repository for the methodology described. |
| Open Datasets | Yes | Datasets. CIFAR10(/100): A natural image dataset containing 50,000/10,000 train/test images from 10 (/100) object classes. Image Net: A large scale 1,000 classes object dataset with 1.2 million images for training and 50,000 for test. SVHN: A Street View House Numbers dataset including 10 classes of digit images. STL10: An Image Net adapted dataset containing 500/800 train/test samples from 10 classes as well as 100,000 unlabelled images from auxiliary unknown classes. MNIST: A hand-written digits dataset with 60,000/10,000 train/test images from 10 digit classes. |
| Dataset Splits | Yes | Datasets. CIFAR10(/100): A natural image dataset containing 50,000/10,000 train/test images from 10 (/100) object classes. Image Net: A large scale 1,000 classes object dataset with 1.2 million images for training and 50,000 for test. SVHN: A Street View House Numbers dataset including 10 classes of digit images. STL10: An Image Net adapted dataset containing 500/800 train/test samples from 10 classes as well as 100,000 unlabelled images from auxiliary unknown classes. MNIST: A hand-written digits dataset with 60,000/10,000 train/test images from 10 digit classes. ... Note that both the training and test datasets are utilised for model learning in the standard clustering setting, unlike the standard classification setting where only the training dataset is used. |
| Hardware Specification | Yes | All the experiments run on Tesla P100 GPU. |
| Software Dependencies | No | The paper mentions software components like 'Alex Net as the backbone', 'SGD with Nesterov momentum', and 'Tarjan framework', but does not specify version numbers for any of these or other key software dependencies. |
| Experiment Setup | Yes | We used Alex Net as the backbone (Krizhevsky, Sutskever, and Hinton 2012). We adopted SGD with Nesterov momentum at 0.9, the epoch number at 200, the initial learning rate at 0.03 with a decay of 0.1 every 40 epochs after the first 80 ones. We set k = 5 (Eq (1)) for graph construction. The maximum size s of SCS is set to 10. We set the weight λ = 0.8 (Eq (5)) and the memory update rate η = 0.5 (Eq (6)). Data augmentation includes horizontal flipping, cropping and colour jittering. |