석사과정생 고동근, 박남준 학생(지도교수: 김재광) 논문 CIKM 2023 Short papers에 채택
- 소셜이노베이션융합전공
- 조회수1231
- 2023-08-10
main Lab.(지도교수: 김재광)의 고동근(인공지능융합학과, 소셜이노베이션융합전공), 이동준(전자전기컴퓨터공학과), 박남준(소프트웨어학과, 소셜이노베이션융합전공-진입예정), 노경래(소프트웨어학과), 박현진(소프트웨어학과) 학생들이 연구한 논문 “AmpliBias: Mitigating Dataset Bias through Bias Amplification in Few-shot Learning for Generative Models” 이 세계 최고 권위 정보검색(Information Retrieval) 및 데이터마이닝 학회인 CIKM 2023 (32nd ACM International Conference on Information and Knowledge Management), Short papers에 채택되었습니다. 논문은 23년 10월 영국 버밍엄에서 발표될 예정입니다.
본 논문은 소셜이노베이션융합전공, 인공지능융합학과와 전자전기컴퓨터공학과 및 소프트웨어학과 석사과정 학생들과 소프트웨어학과 2학년 학부생의 협업을 통한 결과물로서 데이터셋에 존재하는 bias sample로 인한 인공지능 모델의 부정확함을 줄이기 위해 생성모델을 통한 Few shot learning을 하여 Debiased 모델학습 방법을 제안하였습니다. 논문의 자세한 내용은 다음과 같습니다.
[논문]
Donggeun Ko, Dongjun Lee, Namjun Park, Kyoungrae Noh, Hyeonjin Park and Jaekwang Kim, “AmpliBias: Mitigating Dataset Bias through Bias Amplification in Few-shot Learning for Generative Models,” In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management (CIKM 2023), Oct. 2023.
[Abstract].
Deep learning models have demonstrated successful performance in image classification tasks. However, these models exhibit a dependency on peripheral attributes of input data, such as shapes and colors, eventually leading to become biased towards these certain attributes, resulting in subsequent degradation of performance. To address this issue, debiasing techniques have been explored to enhance the robustness of model from biases. Recent debiasing techniques improve biased classifier f_b by reweighting technique or augment the biased dataset to mitigate bias. In this paper, we focus on the latter approach, presenting AmpliBias, a novel framework that tackles dataset bias by leveraging generative models to amplify bias and facilitate the learning of debiased representations of the classifier. Our method involves three major steps. First, we train a biased classifier, f_b, using a biased dataset and extract top-K biased-conflict samples. Subsequently, we train a generator on a bias-conflict dataset composed solely of the top-K samples to learn the distribution of bias-conflict samples. Finally, we re-train the classifier with the new debiased dataset, allowing the biased classifier to competently learn debiased representation. Extensive experiments validate that our proposed method effectively debiases the biased classifier.