Overcoming Language Priors in Visual Question Answering with Adversarial Regularization
Authors: Sainandan Ramakrishnan, Aishwarya Agrawal, Stefan Lee
NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show empirically that it can improve performance significantly on a bias-sensitive split of the VQA dataset for multiple base models achieving state-of-the-art on this task. |
| Researcher Affiliation | Academia | Sainandan Ramakrishnan Aishwarya Agrawal Stefan Lee Georgia Institute of Technology {sainandancv, aishwarya, steflee}@gatech.edu |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper mentions using public codebases for base models (e.g., "SAN Codebase: https://github.com/abhshkdz/neural-vqa-attention"), but does not provide a clear statement or link for the open-sourcing of their own proposed methodology's code. |
| Open Datasets | Yes | We experiment on the VQA-CP dataset [2] with multiple base VQA models, and find 1) our approach provides consistent improvements over all baseline VQA models... We train our models on the VQA-CP [2] train split and evaluate on the test set using the standard VQA evaluation metric [6]. |
| Dataset Splits | Yes | We train our models on the VQA-CP [2] train split and evaluate on the test set using the standard VQA evaluation metric [6]. For each model, we also report results when trained and evaluated on the standard VQA train and validation splits [6, 12] with the same regularization coefficients used for VQA-CP to compare with [2]. |
| Hardware Specification | Yes | The model takes 8 hours to train on a TITAN X for SAN (Torch, 60 epochs) and < 1 hour for Up Down (Py Torch, 40 epochs). |
| Software Dependencies | No | The paper mentions "Torch" and "Py Torch" as frameworks used for training but does not provide specific version numbers for these or any other software dependencies. |
| Experiment Setup | Yes | We set batch size to 150, learning rate to 0.001, weight decay of 0.999 and use the Adam optimizer. |