Semantic Feature Extraction for Generalized Zero-Shot Learning

Authors: Junhan Kim, Kyuhong Shim, Byonghyo Shim1166-1173

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental From extensive experiments using various datasets, we show that the proposed SE-GZSL technique outperforms conventional GZSL approaches by a large margin.
Researcher Affiliation Academia Junhan Kim, Kyuhong Shim, Byonghyo Shim Department of Electrical and Computer Engineering, Seoul National University, Seoul, Korea {junhankim, khshim, bshim}@islab.snu.ac.kr
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any explicit statement or link for open-source code for the methodology.
Open Datasets Yes In our experiments, we evaluate the performance of our model using four benchmark datasets: Aw A1, Aw A2, CUB, and SUN. The Aw A1 and Aw A2 datasets contain 50 classes of animal images annotated with 85 attributes (Lampert, Nickisch, and Harmeling 2009; Xian, Schiele, and Akata 2017). The CUB dataset contains 200 species of bird images annotated with 312 attributes (Welinder et al. 2010). The SUN dataset contains 717 classes of scene images annotated with 102 attributes (Patterson and Hays 2012).
Dataset Splits Yes In dividing the total classes into seen and unseen classes, we adopt the conventional dataset split presented in (Xian, Schiele, and Akata 2017).
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory specifications) used for running its experiments, only mentioning the use of Res Net-101 and MLPs.
Software Dependencies No The paper mentions using Res Net-101 and MLPs, but does not specify software versions for libraries like PyTorch, TensorFlow, or Python, nor specific solvers.
Experiment Setup Yes We set the number of hidden units to 4096 and use Leaky Re LU with a negative slope of 0.02 as a nonlinear activation function. For the output layer of the generator, the Re LU activation is used... The gradient penalty coefficient in the WGAN loss LG,WGAN is set to λgp = 10 as suggested in the original WGAN paper (Gulrajani etm. al. 2017). We set the weighting coefficients in (7), (11), and (16) to λs = 20, λr = 50, λsim = 1, λG,MI = 1, λG,sim = 0.025.