Explanation vs Attention: A Two-Player Game to Obtain Attention for VQA

Authors: Badri Patro, Anupriy, Vinay Namboodiri11848-11855

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the proposed method i.e. PAAN in a number of ways which includes both quantitative analysis and qualitative analysis. Quantitative analysis includes ablation analysis with other variants that we tried using metrics such as Rank correlation (RC) score (Das et al. 2016), Earth Mover Distance (EMD) (Arjovsky, Chintala, and Bottou 2017), and VQA accuracy etc. as shown in table 1 and 2 respectable. We also compare our proposed method with various state of the art models, as provided in table 3 and 4.
Researcher Affiliation Academia Badri N. Patro, Anupriy, Vinay P. Namboodiri Indian Institute of Technology, Kanpur {badri, anupriy, vinaypn}@iitk.ac.in
Pseudocode Yes Algorithm 1 Training PAAN
Open Source Code No The paper states 'We have provided more results... in our project page1. 1https://delta-lab-iitk.github.io/Two Player/', but this does not explicitly state that the source code for the methodology is provided. The linked project page states 'Code will be shared soon'.
Open Datasets Yes Human Attention (HAT) dataset for VQA task; VQA-v1 dataset (Antol et al. 2015); VQA-v2 dataset (Goyal et al. 2017).
Dataset Splits No The paper mentions 'fine-tuned using validation set' and refers to the 'validation set' but does not provide specific details on the split percentages, sample counts, or explicit standard split references for reproduction.
Hardware Specification No The paper does not specify any particular GPU or CPU models, memory, or cloud computing instance details used for experiments.
Software Dependencies No The paper does not mention any specific software dependencies with version numbers (e.g., Python, TensorFlow, PyTorch versions).
Experiment Setup Yes Where n is the number of examples, and η = 10 is the hyperparameter, fine-tuned using validation set and Lc is standard cross entropy loss.