Stratified Rule-Aware Network for Abstract Visual Reasoning

Authors: Sheng Hu, Yuqing Ma, Xianglong Liu, Yanlu Wei, Shihao Bai1567-1574

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments are conducted on both PGM and I-RAVEN datasets, showing that our SRAN outperforms the state-of-the-art models by a considerable margin.
Researcher Affiliation Academia 1 State Key Laboratory of Software Development Environment, Beihang University, Beijing, China 2 Beijing Advanced Innovation Center for Big Data-Based Precision Medicine, Beihang University, Beijing, China husheng 7@163.com, {mayuqing,xlliu}@nlsde.buaa.edu.cn, {weiyanlu,16061167}@buaa.edu.cn
Pseudocode Yes Algorithm 1 Attribute Bisection Tree Input: the correct answer ω 1: Initialize the answer set Ω= {ω } 2: Sample 3 attributes a1, a2, a3 according to ω 3: Sample new value vi for each ai 4: for i = 1 to 3 do 5: Initialize Γ = {} 6: for each wk in the current answer set Ωdo 7: γ modifying attribute ai of ωk with vi 8: Γ Γ S{γ} 9: end for 10: Ω ΩS Γ 11: end for Output: the answer set Ω(|Ω| = 23 = 8)
Open Source Code No The paper does not contain any explicit statements or links indicating that the source code for the described methodology is publicly available.
Open Datasets Yes PGM (Barrett et al. 2018) is another RPM dataset consisting of 1.42M questions.
Dataset Splits No The paper mentions 'train' and 'test' in the context of models and results (e.g., 'we train two models', 'Test on RAVEN and I-RAVEN', 'Test accuracy') but does not specify the explicit percentages or sample counts for training, validation, or test splits, nor does it cite predefined splits with such details.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory amounts used for running its experiments. It only discusses software components and training parameters.
Software Dependencies No The paper mentions several techniques and models (e.g., 'Res Net-18', 'ADAM optimizer', 'dropout') but does not specify any software dependencies with version numbers (e.g., 'PyTorch 1.9' or 'Python 3.8').
Experiment Setup Yes For our SRAN, we adopt three Res Net-18 (He et al. 2016) as the embedding networks for the three hierarchies, by modifying the input channels. The gate fusion ϕ1 and ϕ2 are 2-layer fully connected networks, while ϕ3 is a 4layer fully connected network with dropout (Srivastava et al. 2014) of 0.5 applied on the last layer. We adopt stochastic gradient descent using ADAM (Kingma and Ba 2014) optimizer. The exponential decay rate parameters are β1 = 0.9, β2 = 0.999, ϵ = 10 8. Each reported accuracy is averaged over 5 runs.