15:00-16:30

주제: Transformer 를 활용한 이미지 처리 모델

초록: 지난 10년간 이미지 분류 또는 객체탐지 영역은 Convolutional Neural Networks(CNN) 기반의 딥러닝 모델들이 주로 연구되었고, 이러한 CNN 모델은 현재까지도 여러 산업 분야에서 활용되고 있습니다. 그와 동시 자연어 처리 분야에서 혁신적인 성과를 이끌어낸 Transformer 기술이 최근 비전 영역까지 확대되었고, 많은 연구 결과들을 통해 CNN모델과 비교해서 더 좋은 성능을 얻을 수 있었습니다.

이번 강연에서는 간단한 이미지 분류부터 객체 탐지까지 Transformer 기술이 적용된 딥러닝 모델들을 소개하고, 특히 현재 Object Detection 분야에서 state-of-the-art인 Swin Transformer 모델에 대하여 설명하고자 합니다.

16:30-17:00

토론 및 네트워킹

11:00-13:00

주제: A Novel CT Image Reconstruction Method

초록: We introduce a novel CT image reconstruction method which is also a sinogram recovery method. There are many ways to recover a broken sinogram (X-ray projections). But they use the information of sinogram itself. Unlike other methods, we use the relationship between the tomographic image and the X-ray projections. We also show a comparison result of our method and filtered back projection.

13:00-14:00

토론 및 네트워킹

16:00-17:30

주제: An Algorithm for Peer Reviewer Recommendation Based on Scholarly Activity Assessment

초록: Journal editors are putting a lot of effort into selecting appropriate reviewers for fair and reliable peer review of ted manus. Editors consider whether the reviewers have no affinity with any of the authors of manus and have sufficient expertise in reviewing the manus. The affinity can be evaluated by whether any of the reviewers has been a coauthor and/or a coworker in a common institution with any of the authors of the manu. The expertise depends on the similarity of the research topic between the reviewer’s published papers and the ted manus. In this paper we propose an algorithm to recommend appropriate reviewers to editors, based on the assessment of these scholarly activities and achievements. To implement this algorithm, TextRank and GenSim library are used in order to extract feature sets from abstract and introduction sections of both ted manus and the reviewer candidates’ papers. And then based on the extracted feature sets, affinity and expertise check are implemented. To evaluate the performance of this algorithm, an experiment has been conducted with a data set of over 1,000 papers in the field of DB research. The experiment consists of affinity check by using 2-mode network matrix operations and expertise check based on Max Similarity and/or topic classification. Experimental results show that the recommendation algorithm is reasonable on the basis of scholarly activity assessment.

17:30-18:00

토론 및 네트워킹

14:00-14:50

주제: Lorentzian polynomials (우리말 강연)

• 초록: Lorentzian polynomials link continuous convex analysis and discrete convex analysis via tropical geometry. The tropical connection is used to produce Lorentzian polynomials from discrete convex functions. The talk will be accessible to undergraduate students in mathematics: The audience will need no specific background beyond linear algebra and multivariable calculus to enjoy the presentation. In addition, I advertise the talk and the corresponding paper to people with interests in at least one of the following topics: graphs, convex bodies, stable polynomials, projective varieties, Potts model partition functions, tropicalizations, Schur polynomials, highest weight representations. Based on joint work with Petter Brändén.

14:50-15:00

break

15:00-15:50

토론 및 네트워킹

15:00-16:30

주제: Meta Pseudo Labels & Vision in Transformer

초록: 고전적인 Pseudo labels 방법은 Labeled data로 사전 학습된 teacher model에 Unlabeled data를 넣어 pseudo label을 추출하고, 을 사용하여 student model을 학습하는 방식입니다. 하지만 teacher model이 예측한 pseudo label이 부정확하다면 student model의 성능은 크게 저하되는 문제가 발생합니다. Meta Pseudo Labels은 이러한 Pseudo Labels의 문제를 해결하여 Imagenet benchmark에서 처음으로 90% 이상의 정확도를 달성하였습니다. 이번 발표에서는 Meta Pseudo Labels의 소개와 논문 실험결과 및 한국어 글자 인식에 적용한 실험결과를 소개합니다. 또한 현재 NLP 분야에서 우수한 성능을 자랑하는 Transformer Architecture가 이미지 분야에서는 어떻게 적용되고 있는지 간단히 말씀드리겠습니다.

16:30-17:00

토론 및 네트워킹