Multi-class classification without multi-class labels
Authors: Yen-Chang Hsu, Zhaoyang Lv, Joel Schlosser, Phillip Odom, Zsolt Kira
ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our method is evaluated against state of the art in all three learning paradigms and shows a superior or comparable accuracy, providing evidence that learning multi-class classification without multi-class labels is a viable learning option. |
| Researcher Affiliation | Academia | Yen-Chang Hsu1, Zhaoyang Lv1, Joel Schlosser2, Phillip Odom2, and Zsolt Kira12 1Georgia Institute of Technology 2Georgia Tech Research Institute 1{yenchang.hsu,zhaoyang.lv,zkira}@gatech.edu 2{joel.schlosser,phillip.odom}@gtri.gatech.edu |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The demo code is available at https://github.com/GT-RIPL/L2C |
| Open Datasets | Yes | We compare the classification accuracy on three image datasets: MNIST (Le Cun, 1998) is a 10-class handwritten digit dataset... CIFAR10 and CIFAR100 (Krizhevsky, 2009) instances are colored... Omniglot (Lake et al., 2015): This dataset has 20 images... Image Net (Deng et al., 2009): The 1000-class dataset... |
| Dataset Splits | No | The paper mentions training and testing data sizes (e.g., '60000 images for training, and 10000 for testing' for MNIST) but does not specify a separate validation dataset split for hyperparameter tuning or early stopping. While the semi-supervised learning section discusses labeled and unlabeled portions, this is not a general train/validation/test split for model evaluation. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. |
| Software Dependencies | No | The paper mentions optimizers like 'Adam' and 'SGD' but does not specify any programming languages, libraries, or other software dependencies with version numbers. |
| Experiment Setup | Yes | All networks in this section are trained from scratch with randomly initialized weights. By default, we use Adam... with mini-batch size 100 and initial learning rate 0.001. On MNIST the learning rate was dropped every 10 epochs by a factor of 0.1 with 30 epochs in total. On CIFAR10/100 we use the same setting except that the learning rate is dropped at 80 and 120 epochs with 140 epochs in total. For CIFAR100, the mini-batch size was 1000 and the learning rate dropped at epoch 100 and 150 with 180 epochs in total. In the experiments with Res Net, we use SGD instead of Adam... The learning rate for SGD starts with 0.1 and decays with a factor of 0.1 at the number of epochs described above. |