[한국인공지능학회] Domain-Aware Label Smoothing for Robust Abstract Visual Reasoning - 최승규, 차수빈, 임종민, 김광수
- 인공지능융합 연구실
- 조회수211
- 2023-07-20
Abstract
The Raven’s Progressive Matrices(RPM) prob- lem involves discovering rules within a set of im- ages, serving as an evaluation metric for AI mod- els’ visual reasoning capabilities. Noisy Contrast and Decentralization(NCD) was introduced by Tao et al. to tackle this task. However, we have identified certain limitations with the NCD ap- proach. Specifically, it causes imbalanced label distribution in the preprocessing process, nega- tively affecting the model’s robustness and gen- eralization power. Meanwhile, the panels from other answers could be considered panels in dif- ferent domains. To address these challenges, we propose a novel approach incorporating domain- aware label smoothing. By selectively apply- ing label smoothing to specific rows based on the distinction of domains, we aim to enhance the model’s robustness. Our experimental results on various RPM problem datasets demonstrate the efficacy of domain-aware label smoothing method in improving overall performance and show that the robustness of model calibration improves performance.