A Variational Autoencoder with Deep Embedding Model for Generalized Zero-Shot Learning
Authors: Peirong Ma, Xiao Hu11733-11740
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on four GZSL benchmark datasets show that the proposed model significantly outperforms the state of the arts. |
| Researcher Affiliation | Academia | Peirong Ma, Xiao Hu* School of mechanical and electrical engineering, Guangzhou University, Guangzhou, China mpr_666@163.com, huxiao@gzhu.edu.cn |
| Pseudocode | No | The paper describes the model components and their functions but does not include structured pseudocode or an algorithm block. |
| Open Source Code | No | The paper does not provide an explicit statement about releasing its source code or a link to a code repository. |
| Open Datasets | Yes | The proposed model is evaluated on four widely used ZSL/GZSL benchmark datasets, namely Caltech-UCSDBirds 200-2011 (CUB) (Wah et al. 2010), SUN Attribute (SUN) (Patterson et al. 2012), Animals with Attributes 1 (AWA1) (Lampert et al. 2014) and Animals with Attributes 2 (AWA2) (Xian et al. 2018a). |
| Dataset Splits | Yes | we follow the splits and evaluation protocol proposed by Xian et al. (2018a). In the splits proposed by Xian et al. (2018a), the details of the number of seen and unseen classes images at training and test time are shown in Table 1. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used to run the experiments. |
| Software Dependencies | No | The paper mentions general software components like 'Res Net101' and 'Adam optimizer' but does not specify programming languages, libraries, or their version numbers. |
| Experiment Setup | Yes | The deep embedding network (φ), encoder (E) and decoders (D1 and D2) in the proposed model are implemented as multilayer perceptron (MLP) with one hidden layer. We use 1200 and 1560 hidden units for the deep embedding network and the encoder, respectively. In addition, the D1 and D2 have 1450 and 660 hidden units, respectively. λ is set to 0.9, α and ߛ follow the settings in (Schonfeld et al. 2019). The size of the latent feature is 70 (CUB), 80 (SUN), 65 (AWA1 and AWA2), respectively. The model is trained for 100 epochs using the Adam optimizer, and the batch size is 50. We uses the L2 distance to construct the loss of the deep embedding network, and all other losses use the L1 distance. |