제 1회 아주-인하-고려 산업수학 워크숍

abstract : Systematics efforts for the interaction of mathematics and industry began in 1969 when the first SGI (Study Groups with Industry) was held at Oxford. This talk will briefly survey subsequent activities, and will examine possible steps to expedite the use of mathematics to overcome the challenges that the industry and the society are facing.

abstract : Generation of attributed graphs is important in chemistry, especially in de novo design of organic materials and drugs. For a machine learning model to learn from its generations, it is best to have a loss function that is fast and differentiable. We adopt the kernel distance and propose an approximate solution as a graph similarity. The idea leads to a slight modification of simple graph neural networks which improves both efficiency and prediction quality.

abstract : Distributional shifts in photometry and texture have been extensively studied for unsupervised domain adaptation, but their counterparts in optical distortion have been largely neglected. In this work, we tackle the task of unsupervised domain adaptation for semantic image segmentation where unknown optical distortion exists between source and target images. To this end, we propose a distortion-aware domain adaptation (DaDA) framework that boosts the unsupervised segmentation performance. We first present a relative distortion learning (RDL) approach that is capable of modeling domain shifts in fine-grained geometric deformation based on diffeomorphic transformation. Then, we demonstrate that applying additional global affine transformations to the diffeomorphically transformed source images can further improve the segmentation adaptation. Besides, we find that our distortion-aware adaptation method helps to enhance self-supervised learning by providing higher-quality initial models and pseudo labels. To evaluate, we propose new distortion adaptation benchmarks, where rectilinear source images and fisheye target images are used for unsupervised domain adaptation. Extensive experimental results highlight the effectiveness of our approach over state-of-the-art methods under unknown relative distortion across domains. Datasets and more information are available at https://sait-fdd.github.io/.

abstract : 반도체 산업에서 수요 예측은 Pricing, Production Planning, Biz Simulation 등 중요 의사결정의 기반이 되는 핵심 경영관리 기능이다. 특히 다양한 외부 요인과 대규모 설비 변경이 필요해 빠른 대응이 어려워 중장기 예측이 중요하다. 하지만, 반도체 시장의 불확실성과 변동성이 커서 예측이 어려워 최근 정확도를 향상시키기 위해 뉴스 기사, 증권사/투자기관 리포트 등의 비정형 데이터 활용 필요성이 대두되고 있다. 본 워크숍에서는 수요 예측을 위한 비정형 데이터 기반 인공지능 기술로 Knowledge Graph를 활용하는 Event 중심의 마켓 시황 분석 및 예측 방법을 소개한다. 상세 기술 토픽은 multi-modality graph neural network, event extraction and prediction 을 통한 태스크 정의와 모델링이다.

abstract : In this talk, we show results in anomaly detection algorithms for an oral health care application. In particular, one class YOLOv3 is investigated as an anomaly detection model to classify pictures of mouths which will be used as inputs in the following machine learning model. We have achieved outstanding performances by proposing appropriate annotation strategies for our data sets and modifying the loss function. Indeed, in the model with the best performance, classification evaluation metrics such as accuracy, precision, recall, F1-score and AUC are greater than 0.99. Moreover, the model can not only classify oral and non-oral pictures, but also output preprocessed pictures that only contain the area around the lips by using the predicted bounding box.

abstract : AI & Math Lab at Korea Univ. Department of Math presents a virtual fashion garment wearing algorithm. The algorithm transforms flat-on-floor garment photographs into avatar-form-fitted garment. Moreover, the algorithm can produce 3D version of the garment as well. This work is yet another evidence of the potential of machine learning in reconstructing functions projected to a lower dimension back into a higher dimensional space with artificially added regularizations.