Self-Critical Reasoning for Robust Visual Question Answering

Authors: Jialin Wu, Raymond Mooney

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our approach on the VQA generalization task using the VQA-CP dataset, achieving a new state-of-the-art i.e., 49.5% using textual explanations and 48.5% using automatically annotated regions.
Researcher Affiliation Academia Jialin Wu Department of Computer Science University of Texas at Austin jialinwu@utexas.edu Raymond J. Mooney Department of Computer Science University of Texas at Austin mooney@cs.utexas.edu
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https://github.com/jialinwu17/Self_Critical_VQA.
Open Datasets Yes We evaluate our approach on the VQA generalization task using the VQA-CP dataset... We also report our system s performance on the balanced VQA v2 validation set for completeness. The Expl. column shows the source of explanations for training the VQA systems.
Dataset Splits Yes We first pre-train our base Up Dn VQA system on the VQA-CP training set using standard VQA loss Lvqa (binary cross-entropy loss with soft scores as supervision) with the Adam optimizer [16] for at most 20 epochs... Then, we fine-tune our system to recognize important objects using Lvqa + λinfl Linfl with a learning rate of 10e-5 for at most 15 epochs on the intersection of VQA-X and VQA-CP training set.
Hardware Specification No The paper does not explicitly describe the hardware used for running its experiments.
Software Dependencies No The paper mentions software tools and libraries like Adam optimizer, GRU, Glove embeddings, Faster R-CNN, ResNet-101, and spaCy POS tagger, but does not provide specific version numbers for these dependencies.
Experiment Setup Yes We first pre-train our base Up Dn VQA system on the VQA-CP training set using standard VQA loss Lvqa (binary cross-entropy loss with soft scores as supervision) with the Adam optimizer [16] for at most 20 epochs. As suggested in [27], the learning rate is fixed to 10e-3 with a batch size of 384 during the pre-training process, and we use 1, 280 hidden units in the base Up Dn VQA system. Then, we fine-tune our system to recognize important objects using Lvqa + λinfl Linfl with a learning rate of 10e-5 for at most 15 epochs... Finally, we fine-tune the system with the joint loss L = Lvqa + λ infl Linfl + λcrit Lcrit for at most 15 epochs with a learning rate of 10e-5... The bucket size |B| of the competitive answers is set to 5...