Semantic-guided Reinforced Region Embedding for Generalized Zero-Shot Learning

Authors: Jiannan Ge, Hongtao Xie, Shaobo Min, Yongdong Zhang1406-1414

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on four public benchmarks demonstrate that the proposed SR2E is an effective GZSL method with reinforced embedding space, which obtains averaged 6.1% improvements.
Researcher Affiliation Academia Jiannan Ge, Hongtao Xie , Shaobo Min, Yongdong Zhang School of Information Science and Technology, University of Science and Technology of China, Hefei, China {gejn, mbobo}@mail.ustc.edu.cn, {htxie, zhyd73}@ustc.edu.cn
Pseudocode No The paper describes the methods in text but does not include structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any explicit statement or link regarding the public availability of its source code.
Open Datasets Yes We evaluate the proposed SR2E model on four GZSL benchmarks: Caltech-UCSD Birds 200-2011(CUB) (Wah et al. 2011), Animals with Attributes 2(Aw A2) (Lampert, Nickisch, and Harmeling 2013), SUN (Xiao et al. 2010) and Yahoo dataset(a PY) (Farhadi et al. 2009).
Dataset Splits Yes We adopt the splits of seen/unseen classes proposed in (Xian et al. 2018a). ... We divide the validation set from the training set to estimate the threshold τ. ... We follow the settings in (Xian et al. 2018a) to adopt Mean Class Accuracy (MCA) as the evaluation indicator.
Hardware Specification No We also acknowledge the support of GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC.
Software Dependencies No The paper mentions using ResNet101 and Adam optimizer, but does not specify any software libraries or frameworks with version numbers (e.g., PyTorch, TensorFlow, Python versions).
Experiment Setup Yes Regarding the parameters, we set learning rate lr = 1 × 10−4 and adopt Adam optimizer with β = (0.5, 0.999) and weigh decay 1 × 10−4 in the first stage. In the second stage, the learning rate of the action predictor is fixed and set to 1 × 10−6, while the termination action reward value η is set to 3. We set α = 6/7 for the animal datasets like CUB and Aw A2 because their images have specific objects which are useful for classification. SUN and a PY datasets have scene pictures, which means the most areas of the image affect prediction. In this case, we set α = 9/10. The maximum number of iterations is set to 6 to avoid invalid search. The ϵ value in ϵ greedy is initialized to 1, and decreases by 0.1 at each epoch until the 10th epoch.