Rethinking Data Augmentation for Robust Visual Question Answering

Abstract

Being widely used in learning unbiased visual question answering (VQA) models, Data Augmentation (DA) helps mitigate language biases by generating extra training samples beyond the original samples. While today's DA methods can generate robust samples, the augmented training set, significantly larger than the original dataset, often exhibits redundancy in terms of difficulty or content repetition, leading to inefficient model training and even compromising the model performance. To this end, we design an Effective Curriculum Learning strategy ECL to enhance DA-based VQA methods. Intuitively, ECL trains VQA models on relatively "easy" samples first, and then gradually changes to "harder" samples, and less-valuable samples are dynamically removed. Compared to training on the entire augmented dataset, our ECL strategy can further enhance VQA models' performance with fewer training samples. Extensive ablations have demonstrated the effectiveness of ECL on various methods.

Remote sensing and machine learning to improve aerial wildlife population surveys

“…Improving the accuracy of machine learning models to reduce reliance on human review in terms of annotating inputs and in correcting errors will also be critical. This will necessitate foundational computer vision research into avenues such as active learning techniques (e.g., Norouzzadeh et al, 2021), data augmentations (e.g., Chen et al, 2022), new model architectures, and more consistent image processing procedures (e.g., Kellenberger et al, 2021). Increasing confidence in machine learning outputs through improved statistical rigor will be necessary, as current methods do not adequately characterize multiple sources of error and bias including the downstream propagation of uncertainty in image labels.…”

Semi-Supervised Implicit Augmentation for Data-Scarce VQA

“…Additionally, generative models are computationally expensive to train and require large amounts of data to produce good-quality results. Some works that are not generation based are SimpleAug [25] and KDDAug [26]. SimpleAug works on the principle that many of the "unknowns" are indeed "known" within the dataset.…”