Exploring and exploiting model uncertainty for robust visual question answering

Visual Question Answering (VQA) methods have been widely demonstrated to exhibit bias in answering questions due to the distribution differences of answer samples between training and testing, resulting in resultant performance degradation. While numerous efforts have demonstrated promising results in overcoming language bias, broader implications (e.g., the trustworthiness of current VQA model predictions) of the problem remain unexplored. In this paper, we aim to provide a different viewpoint on the problem from the perspective of model uncertainty. In a series of empirical studies on the VQA-CP v2 dataset, we find that current VQA models are often biased towards making obviously incorrect answers with high confidence, i.e., being overconfident, which indicates high uncertainty. In light of this observation, we: (1) design a novel metric for monitoring model overconfidence, and (2) propose a model calibration method to address the overconfidence issue, thereby making the model more reliable and better at generalization. The calibration method explicitly imposes constraints on model predictions to make the model less confident during training. It has the advantage of being model-agnostic and computationally efficient. Experiments demonstrate that VQA approaches exhibiting overconfidence are usually negatively impacted in terms of generalization, and fortunately their performance and trustworthiness can be boosted by the adoption of our calibration method. Code is available at https://github.com/HCI-LMC/VQA-Uncertainty © The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2024.