Adaptive and Generative Zero-Shot Learning

Authors: Yu-Ying Chou, Hsuan-Tien Lin, Tyng-Luh Liu

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

Reproducibility Variable Result LLM Response
Research Type Experimental To demonstrate the effectiveness of our method, we report state-of-the-art results on four standard GZSL datasets, including an ablation study of the proposed modules.
Researcher Affiliation Collaboration Yu-Ying Chou1,3, Hsuan-Tien Lin3 & Tyng-Luh Liu1,2 1Institute of Information Science, Academia Sinica, Taiwan 2Taiwan AI Labs, Taiwan 3Department of Computer Science, National Taiwan University, Taiwan
Pseudocode No The paper describes the methodology in detail and includes architectural diagrams (e.g., Figure 1, Figure 2, Figure 3), but it does not contain any formally labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any statements about releasing open-source code or provide a link to a code repository for the described methodology.
Open Datasets Yes Datasets. In GZSL, there are four widely used benchmark datasets for evaluation, which are AWA2 (Xian et al., 2018a), USCD Birds-200-2011 (CUB) (Welinder et al., 2010), SUN (Patterson & Hays, 2012), and a PY (Farhadi et al., 2009).
Dataset Splits Yes We then split the data into seen and unseen classes according to the benchmark procedure from Xian et al. (2017).
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., GPU models, CPU types, memory) used for running the experiments.
Software Dependencies No The paper mentions software components like 'Res Net101 backbone', 'Adam optimizer', 'Re LU', and 'softmax', but it does not specify version numbers for these or for broader software dependencies like Python, PyTorch, or TensorFlow.
Experiment Setup Yes Implementation. As suggested in Li et al. (2019b), we normalize the visual and semantic features into [0, 1]. The architecture of semantic-to-visual embedding contains a two-layer linear model with 1,600 hidden units and utilizes Re LU on the hidden and output layer. The seen and unseen experts are trained by Adam optimizer with a learning rate 5 10 5 and 5 10 4 respectively for all datasets. We apply 200,000 episodes to train the unseen expert. In each episode, we randomly generate 16 or 20 (based on the dataset) virtual classes and 4 samples for each class.