Learning Classifiers for Target Domain with Limited or No Labels
Authors: Pengkai Zhu, Hanxiao Wang, Venkatesh Saligrama
ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on benchmark datasets. Our method outperforms state-of-art methods trained for the specific contexts (ZSL, FSL, DA). |
| Researcher Affiliation | Academia | 1Electrical and Computer Engineering Department, Boston University, Boston, USA. Correspondence to: Venkatesh Saligrama <srv@bu.edu>. |
| Pseudocode | No | The paper describes its models and algorithms using text, mathematical equations, and network diagrams (e.g., Figure 1), but it does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statements about releasing source code or provide a link to a code repository for the methodology described. |
| Open Datasets | Yes | Datasets. We first evaluate the few shot learning performance of the proposed model on two benchmark datasets: Omniglot(Lake et al., 2015) and mini Image Net(Vinyals et al., 2016)... The performance of our model for GZSL is evaluated on three commonly used benchmark datasets: Caltech UCSD Birds-200-2011 (CUB) (Wah et al., 2011), Animals with Attributes 2 (AWA2) (Xian et al., 2018a) and Attribute Pascal and Yahoo (a PY) (Farhadi et al., 2009)... We evaluate our proposed model in unsupervised domain adaptation task between three digits datasets: MNIST(Le Cun et al., 1998), USPS and SVHN(Netzer et al., 2011). |
| Dataset Splits | Yes | For mini Image Net, the dataset is split into 64 training, 16 validation and 20 testing classes. As such the validation set is used primarily to validate training performance. |
| Hardware Specification | No | The paper mentions training models and parameters but does not provide specific details about the hardware used (e.g., GPU/CPU models, memory specifications) for running the experiments. |
| Software Dependencies | No | The paper mentions using 'Adam optimizer' but does not specify any software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | We set the number of parts M to 4 and in each part the number of prototypes K is set to 16. ζ in Eq.(5) is empirically set to 0.02. For FSL, we set the input size to be [224 224], and λ in Eq.(3) is 2; for GZSL, our model takes input image size as [448 448] and λ is set to 5; For DA, the input image size is [224 224] and λ is set to 2. The task-specific predictor V ( ) for both GZSL and DA is implemented by a two FC-layer neural network with Re LU activation, the number of neurons in the hidden layer is set to 32. Our model is trained for 80 and 30 epochs on Omniglot and mini Image Net, repectively. The learning rate for step.A is set to 1e-6, and the learning rate of step.B is 1e-4 for Omniglot and 1e-5 for mini Image Net. |