Task Aligned Generative Meta-learning for Zero-shot Learning

Authors: Zhe Liu, Yun Li, Lina Yao, Xianzhi Wang, Guodong Long8723-8731

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments show TGMZ achieves a relative improvement of 2.1%, 3.0%, 2.5%, and 7.6% over state-of-the-art algorithms on AWA1, AWA2, CUB, and a PY datasets, respectively. Overall, TGMZ outperforms competitors by 3.6% in the generalized zero-shot learning (GZSL) setting and 7.9% in our proposed fusion ZSL setting. We conduct extensive experiments on four benchmark datasets: AWA1 (Lampert, Nickisch, and Harmeling 2009), AWA2 (Xian et al. 2019), CUB (Welinder et al. 2010), and a PY (Farhadi et al. 2009).
Researcher Affiliation Academia Zhe Liu,1* Yun Li, 1 Lina Yao, 1 Xianzhi Wang, 2 Guodong Long 2 1 University of New South Wales, Australia 2 University of Technology Sydney, Australia {zhe.liu1, yun.li5, lina.yao}@unsw.edu.au, {xianzhi.wang, guodong.long}@uts.edu.au
Pseudocode Yes Algorithm 1 TGMZ Training Procedure
Open Source Code No The paper does not provide any concrete access to source code (no specific repository link, explicit code release statement, or mention of code in supplementary materials).
Open Datasets Yes We conduct extensive experiments on four benchmark datasets: AWA1 (Lampert, Nickisch, and Harmeling 2009), AWA2 (Xian et al. 2019), CUB (Welinder et al. 2010), and a PY (Farhadi et al. 2009).
Dataset Splits Yes Different from conventional ZSL settings, we divide the training set Dtr into a support set Dsup and a disjoint query set Dqry to mimic seen and unseen classes, respectively, during training. We divide the datasets into seen and unseen classes following the proposed split (PS) (Xian et al. 2019) and adopt visual features from pre-trained Res Net-101, according to Xian et al. (Xian et al. 2019). The dataset statistics and train/test split are shown in Table 1.
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types, or memory amounts) used for running its experiments.
Software Dependencies No The paper mentions using 'pre-trained Res Net-101' and evaluating with 'Linear-SVM' and 'Softmax' but does not specify any software names with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes By default, we set σ = 1 and the sample number to 100. More details about Dataset Description, Model Architecture, Parameter Setting and Convergence Analysis can be found in Supplementary Material.