http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
비대칭 데이터 인식을 위한 배치 단위 확률 추정 및 학습률 분리
손병관(Byunggwan Son),함범섭(Bumsub Ham) 대한전자공학회 2023 대한전자공학회 학술대회 Vol.2023 No.6
The real world data is often long-tail distributed with highly uneven number of samples per each class. When it comes to DNNs, the networks are skewed towards a few dominating classes when trained naively. Many research efforts were taken to address this problem, but they are still struggling with evenly distributed test images. In this paper, we propose two simple yet effective techniques to improve performance of the long-tail recognition on top of previously proposed research works. First, we propose to estimate skewed probability in batch-level instead of whole training data for logit adjustment method. Second, inspired by the work of decoupling feature extractor and classifier, we propose to decouple their learning rate for more efficient learning. We empirically show that our techniques are outperforming the baseline methods.