Active Generative Adversarial Network for Image Classification
Authors: Quan Kong, Bin Tong, Martin Klinkigt, Yuki Watanabe, Naoto Akira, Tomokazu Murakami4090-4097
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | With extensive evaluations, we have confirmed the effectiveness of the model, showing that the generated samples are capable of improving the classification performance in popular image classification tasks. |
| Researcher Affiliation | Industry | Quan Kong, Bin Tong, Martin Klinkigt, Yuki Watanabe, Naoto Akira, Tomokazu Murakami Hitachi, Ltd. R&D Group, Japan Hitachi America, Ltd. USA {quan.kong.xz, bin.tong.hh, martin.klinkigt.ut}@hitachi.com {naoto.akira.vu, tomokazu.murakami.xr}@hitachi.com yuki.watanabe@hal.hitachi.com |
| Pseudocode | Yes | Algorithm 1 Active GAN Input training data xi and its label yi where i [1, . . . , N]. Output Ψd (parameters of D), Ψg (parameters of G) and θ (parameters of MLP) 1: Initialize α, λ, θ, Ψd and Ψg. 2: Set the buffer size to be M 3: Train SVM with grid-search for best parameters 4: Train the generator G and the discriminator D with first m iterations 5: Save generated samples in m iterations into the buffer 6: repeat 7: Generate a sample bxi G(z, yi) 8: Use Equation 9 to calculate the reward r(bxi) for bxi 9: Use generated samples to calculate the likelihood P(bxi|θ) for bxi 10: Use Equation 10 to calculate the loss LU related to the degree of uncertainty for bxi 11: Update parameters for the generator G and MLP: Ψg,θ (Ψg,θ) + Ψg,θ LG Active GAN(Ψg, θ) 12: Update parameters for the discriminator D: Ψd Ψd + Ψd LD AC-GAN 13: Update the buffer by adding the sample bxi 14: until |
| Open Source Code | No | The paper does not provide any explicit statement or link indicating the release of source code for the described methodology. |
| Open Datasets | Yes | We utilized four datasets CIFAR10 (Krizhevsky, Nair, and Hinton ), MNIST (Netzer et al. 2011), Fashion-MNIST (Xiao, Rasul, and Vollgraf 2017) and a large scale dataset Tiny-Image Net (Russakovsky et al. 2015) for evaluation of the proposed model Active GAN. |
| Dataset Splits | Yes | MNIST consists of 50,000 training samples, 10,000 validation samples and 10,000 testing samples of handwritten digits of size 28 28. CIFAR10 has colored images for 10 general classes. Again we find 50,000 training samples and 10,000 testing samples of size 32 32 in CIFAR10. Fashion-MNIST has a training set of 60,000 examples and a test set of 10,000 examples. [...] Tiny-Image Net has 200 classes, each class has 500 training images, 50 validation images, and 50 test images. |
| Hardware Specification | No | The paper does not provide any specific details regarding the hardware (e.g., GPU/CPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions software like 'Adam' and 'SVM' and 'VGG-16' but does not provide specific version numbers for any of them. |
| Experiment Setup | Yes | Adam was used as the gradient method for learning parameters of the network. Its initial learning rate is searched in the set {0.0002, 0.001}. We used SVM as a base classifier and its optimal hyper-parameters are chosen via a grid search. We used a pre-trained VGG-16 (Simonyan and Zisserman 2014) to extract features for images for all datasets. The threshold ϵ in Equation 8 was set to 0.2. The balancing parameter α in Equation 9 was set to 0.5. The balancing parameter λ in Equation 11 was set to 0.1 to guarantee that values of two terms LG AC-GAN and LU are in the same scale. |