Leverage Class-Specific Accuracy to Guide Data Generation for Improving Image Classification
Authors: Jay Gala, Pengtao Xie
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on both imbalanced and balanced classification datasets demonstrate the effectiveness of our proposed method. |
| Researcher Affiliation | Academia | 1University of California San Diego 2Mohamed bin Zayed University of Artificial Intelligence. |
| Pseudocode | Yes | Algorithm 1 Optimization algorithm |
| Open Source Code | No | The paper does not contain any explicit statement about releasing source code or a link to a code repository for the methodology described. |
| Open Datasets | Yes | We used three class-imbalanced datasets: CIFAR10-LT (Cui et al., 2019b), CIFAR100-LT (Cui et al., 2019b), and Image Net-LT (Liu et al., 2019b), where LT denotes long tail. They were curated from the original balanced CIFAR-10 (Krizhevsky et al., 2009), CIFAR100 (Krizhevsky & Hinton, 2010), and Image Net (Deng et al., 2009) |
| Dataset Splits | Yes | Let D(tr) cls = {(xi, yi)}N i=1 denote the training set of an image classification dataset and D(val) cls denote a validation set. From the original 1.3M training images, we randomly sample 10% as training data and another 2.5% as validation data, to perform architecture search. |
| Hardware Specification | Yes | We conducted all experiments on Nvidia 1080Ti GPUs. |
| Software Dependencies | No | The paper mentions optimizers like Adam and SGD but does not provide specific version numbers for these or other key software components, nor does it list the ML framework used with its version (e.g., PyTorch 1.9). |
| Experiment Setup | Yes | The tradeoff parameters λ and γ in Eq.(7) were set to 1. We tuned them in {0.1, 0.5, 1, 2, 3} on 5K held-out examples. The number of epochs was set to 100. The initial learning rate was set to 0.001, which was reduced by 10 after 80 and 90 epochs. Adam (Kingma & Ba, 2014b) was used as the optimizer. Batch size was set to 64. |