Implicit Non-linear Similarity Scoring for Recognizing Unseen Classes
Authors: Yuchen Guo, Guiguang Ding, Jungong Han, Sicheng Zhao, Bin Wang
IJCAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on several benchmark datasets. The experimental results demonstrate consistent accuracy improvement over existing approaches for (generalized) ZSL, which validates the effectiveness of ICINESS. |
| Researcher Affiliation | Academia | Yuchen Guo , Guiguang Ding , Jungong Han , Sicheng Zhao , Bin Wang School of Software, Tsinghua University, Beijing 100084, China School of Computing & Communications, Lancaster University, UK |
| Pseudocode | No | The paper describes the architecture and training process of the network, but it does not include any formal pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any statement or link indicating that the source code for the described methodology is publicly available. |
| Open Datasets | Yes | Following [Xian et al., 2017], we adopt 4 benchmark datasets for ZSL. The first is Animals with Attributes2 (Aw A2) [Xian et al., 2017] with 40 seen classes and 10 unseen classes. The second is a Pascal-a Yahoo (a PY) [Farhadi et al., 2009] with 20 seen classes and 12 unseen classes. The third is SUN [Patterson and Hays, 2012] scene recognition dataset with 645 seen classes and 72 unseen classes. The last is CUB [Wah et al., 2011] bird recognition dataset with 150 seen classes and 50 unseen classes. ... The seen-unseen split is from Xian et al. [2017]. |
| Dataset Splits | Yes | We use all samples from the seen classes, including both train and test in Table 1, to train the network. ... In GZSL, the test data contains two parts. One is the unseen samples and the other is the test seen samples (the fourth row in Table 1). We only use the train seen samples (the third row in Table 1) as the training set. ... The seen-unseen split is from Xian et al. [2017]. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments, such as CPU or GPU models. |
| Software Dependencies | No | The paper mentions 'Tensor Flow framework' but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | Mini-batch based stochastic gradient descent is used to optimize network parameters with initial learning rate 0.01 and decrease it to 0.001 after 75k batches. The training stops after 100k batches. We use 128 image-class pair in each batch. In particular, a positive pair is given by an image and its ground truth class label. A negative pair is an image and a randomly sampled other class. In this paper, we set the positive:negative ratio to 1 : 3, i.e., for each positive class, we randomly sample 3 negative classes for an image in a batch. Therefore, each batch has 32 images and each image has 1 positive class and 3 negative classes, leading to 128 pairs in each batch. The weight decay is 5e-5 and we use random flip and random rotate for data augmentation. |