KCC 2022 여성위원회 특별 세션
신진 여성 연구자 발표세션: 여성연구인 Meet Up Day
여성과학자간 네트워킹 강화 및 박사학위 졸업을 앞두었거나, 학위 취득 후 얼마 지나지 않은 국내외 신진 여성연구인력의 연구성과를 소개하고 기존 여성연구인들과의 밋업데이를 통해 신규 여성리더를 발굴하고 교수진 및 기관과의 네트워킹을 통해 새로운 기회의 장을 마련하고자 한다.
일시: 6.30(목) 09:30-12:00
장소: ICC제주 301A호
프로그램
09:30 ~ 09:45 인사 및 참석자 소개
모든 참석자
09:45 ~ 10:00 Multimodal data integration and machine learning for graphs
김소연 교수(아주대)
10:00 ~ 10:15 Large-scale Image and Video Understanding using Symbolic Graphs
김은솔 교수(한양대)
10:15 ~ 10:30 Formal language theoretical perspective in natural computing
조다정 교수(아주대)
10:30 ~ 10:45 AI for Good: Security, Privacy and Fairness
박새롬 교수(성신여대)
10:45 ~ 11:00 Break
11:00 ~ 11:15 한글 글꼴 형태소 분석 기반 폰트 추천 기술 개발
송유정 교수(세명대)
11:15 ~ 11:30 Rich and Intuitive Haptic Interaction for Future Computers?
이재연 교수(UNIST)
11:30 ~ 11:45 Hypergraph Transformer: Weakly-supervised Multi-hop Reasoning for Knowledge-based Visual Question Answering
허유정 연구원(서울대)
11:45 ~ 12:00 Closing
The real-world data are often represented as relational data such as social relationships, biological pathways, protein interactions. To understand the complex relationships between objects, it is fundamental to design a tool for modeling such graph structured data. In addition, the information in the real world usually comes as different modalities. For example, social-tagged images or multi-omics data in biomedical field. Learning graph structured data or multi-modal data is challenging task due to their complexity and data heterogeneity. This talk introduces several approaches of multi-modal data integration and machine learning techniques for graphs and the current research directions towards graph representation learning.
In this talk, symbolic graphs are introduced as a new framework to understand the large-scale multimodal dataset, such as images and videos. We postulate the disparity of information levels between modalities as a fundamental problem in multimodal learning. To tackle this problem, we propose a new method which combines the multimodal inputs on symbolic levels, which is called Hypergraph Attention Networks (HANs). The main idea of the HANs is to abstract symbolic information from each modality in graph forms and to integrate the symbolic graphs into a coherent concept using graph matching methods.
The aim of the talk is to introduce a relationship between formal language theory and the study of molecule based computation. Formal language theory is a branch of theoretical computer science that is devoted to the study of sets of finite strings of symbols. Since gene sequences such as DNAs, RNAs or protein sequences can be regarded as strings over 4 symbols {A, G, C, T (R)}, where A denotes Adenine, C denotes Cytosine, T denotes Thymine (U denotes Uracil for RNA) and G denotes Guanine, formal language captures their behaviors and analyze the computational properties. From a formal language viewpoint, our research focusses on (1) modeling abnormal gene arrangements such as insertion, deletion, substitution, (2) modeling laboratory techniques that produce rearrangements at specific sites of gene sequences, under enzymatic activities, (3) simulating molecular process such as RNA folding, cellular communication, and (4) analyzing theoretical properties of molecular computation. The goal is to understand computational properties of such gene rearrangements and to identify efficient ways to carry out molecular tasks, via their formal language and automata theory models. This approach could lead to a fruitful link between theoretical and experimental work with DNA sequences and, more generally, between theoretical bio-computing and molecular biology.
As the data-driven industry develops and diversifies, concerns have been raised about the impact of the use of AI. For example, there are possible attack surfaces to hinder the training of AI (i.e., poisoning attacks), breach the confidential data (i.e., model information, private data), and degrade the performance of AI (i.e., evasion attacks). In addition, algorithmic discrimination is one of the significant concerns in applying the AI algorithms to a real-world system. Thus, in this presentation, I want to discuss how these aspects affect the use of AI and introduce my recent studies to address them.
본 연구에서는 문서작성 및 디자인 개발 시 폰트 선택에 있어서 폰트 전문가가 아닌 일반인이 주어진 문장이나 이미지의 분위기에 어울리는 폰트의 유연한 선택을 위한‘글꼴 유사성 판단 기반 어울림 폰트 추천 기술’을 개발한다. 최근 폰트 추천 알고리즘의 개발 동향은 주로 이미지 변환 방식의 딥러닝 기술을 활용하고 있으며, 단순히 이미지 단위 비교를 수행하는 방식으로 이루어지기 때문에 글자의 구조정보를 기반으로 하는 폰트 고유 특성의 반영이 부족하고, 결과적으로 폰트 추천 알고리즘 및 시스템의 추천 품질이 만족스럽지 못하다. 따라서, 단순 이미지 변환 방식의 딥러닝이 아닌, 글꼴의 구조정보에 기초하는 딥러닝 기법의 연구를 통해 다양한 상황에 따라 모양-의미-감정이 어울리는 폰트 추천 기법이 필요하며, 이를 위해 학제간의 융합을 통해 디자인적 감각과 공학적 지식이 융합 된 고품질 폰트 추천 시스템을 구축하는 것이 본 연구의 최종 목표이다.
Computers became small yet powerful enough to be worn and to provide information to the user in daily life. However, interacting with those computers is still challenging, primarily due to their small and rigid form factors. This is problematic since one of the significant reasons for wearing computers is to access information from anywhere in a comfortable way. This talk introduces studies enriching expressivity and natural interactions on small computers using the human sense of touch.
Knowledge-based visual question answering (QA) aims to answer a question which requires visually-grounded external knowledge beyond image content itself. Answering complex questions that require multi-hop reasoning under weak supervision is considered as a challenging problem since i) no supervision is given to the reasoning process and ii) high-order semantics of multi-hop knowledge facts need to be captured. In this talk, we introduce a concept of hypergraph to encode high-level semantics of a question and a knowledge base, and to learn high-order associations between them. The proposed model, Hypergraph Transformer, constructs a question hypergraph and a query-aware knowledge hypergraph, and infers an answer by encoding inter-associations between two hypergraphs and intra-associations in both hypergraph itself. Extensive experiments on two knowledge-based visual QA and two knowledge-based textual QA demonstrate the effectiveness of our method, especially for multi-hop reasoning problem
[최종 수정일: 2022.07.05]