Invited Talk 1.
금융 및 은행에서의 설명가능한 AI(Explainable AI (XAI) in Finance and Banking industry)
오순영 (과실연)
설명 가능성의 개념은 금융 부문에서 특히 중요하다. 이는 금융 부문이 다른 많은 산업보다 더 높은 사회적 기준을 따르기 때문이며, AI를 통한 결정은 금융 시스템의 평판에 중요한 영향을 미칠 수 있기 때문이다. 금융 기관의 경우 관련 프로세스를 설명할 필요성이 있기 때문에 거버넌스 역시 함께 고려되어야 한다. XAI는 은행 운영에 가치를 더하는 동시에 편향 가능성을 해결하고 행동, 규정 준수, 규제 및 감사와 관련된 리스크를 줄일 수 있는 분명한 통찰력을 제공할 것이며, 은행의 미래를 위한 중요한 도구로서 모든 이해관계자 간의 투명성과 신뢰를 향상시킬 것이다.
The concept of explainability is particularly crucial in the financial sector. This is because the financial industry adheres to higher societal standards compared to many other industries, and AI decision-making can significantly impact the reputation of financial systems. Given the necessity for financial institutions to explain their processes, governance must also be considered. Explainable AI (XAI) not only adds value to banking operations but also addresses bias potential, provides clear insights into areas where it can mitigate risks related to behavior, regulatory compliance, and audit, thus enhancing transparency and trust among all stakeholders for the future of banking.
Invited Talk 2.
멀티모달 대화형 인공지능 관련 연구 동향(Towards Multimodal Conversational AI)
서홍석 교수 (고려대학교 정보대학 컴퓨터학과)
우리 인간은 멀티모달 세상을 살아가며 여러 감각 체계를 통해 이같은 멀티모달 세상을 인지하고 이해한다. 한편, 자연어는 오랜시간 사람 간의 커뮤니케이션의 주요 매체로 사용되어 왔다. 멀티모달 대화형 AI 에이전트는 이 둘, 즉 지각과 커뮤니케이션을 연결하기 위해 멀티모달 입력 신호로부터 주어진 상황과 문맥을 이해하고 이를 기반으로 인간과 자연스럽게 대화하는 지능을 가진 에이전트를 의미한다. 이러한 멀티모달 대화형 AI 에이전트를 구축하는 것은 인공 일반 지능을 실현하기 위한 주요 단계 중 하나로 본 강연에서는 최근의 멀티모달 대화형 AI와 관련 된 최신의 연구들을 소개하고 앞으로의 연구 방향과 다른 분야의 연구자들과의 잠재적인 협업에 대해 논의한다.
Humans perceive the world through multiple sensory systems (ex., vision, audition, touch, smell), which work together complementing each other and therefore it is rather natural to build a model processing multiple modalities simultaneously. Based on this understanding of multimodal data, AI systems should communicate with human users through language, which is the main medium for human communication. In this talk, I will introduce our endeavors that expand the realm of multimodal conversational AI where multimodal understanding of the world and communication in language are connected.
Invited Talk 3.
Multimodal Counterfactual reasoning을 이용한 자연어 해석 연구 동향(Natural Language Interpretation Study using Multimodal Counterfactual Reasoning)
박천음 교수 (국립한밭대학교 컴퓨터공학과)
Counterfactual reasoning은 인간 지능의 중요한 표현으로서, 확립된 사실을 바탕으로 전제를 세우고 잠재적인 결과를 추론하는 것을 의미한다. 최근, counterfactual reasoning을 평가하기 위한 벤치마크를 구성하여 Large Language Model (LLM)과 Multimodal LLM (MLLM)을 이용한 연구가 많이 진행됐다. 본 강연에서는 counterfactual reasoning 을 이용한 자연어 해석 연구 동향에 대하여 살펴본다.
Counterfactual reasoning is a significant aspect of human intelligence, involving the formulation of premises based on established facts and the inference of potential outcomes. Recently, a substantial amount of research utilizing Large Language Models (LLMs) and Multimodal LLMs (MLLMs) has been conducted, aimed at developing benchmarks for evaluating counterfactual reasoning. This presentation will explore the trends in natural language explanation research that employ counterfactual reasoning.
Invited Talk 4.
Digital Health에서의 XAI의 역할: Repeech 사례를 중심으로(What can XAI do for digital health? focusing on stroke rehabilitation DTx)
김진우 교수/CEO (연세대학교 경영대학/HAII)
의료 인공지능에서의 오류 발생이 치명적인 결과로 이어질 수 있는 분야인만큼 최근 이를 해결하기 위해 설명가능한 인공지능의 도입이 활발하게 이루어지고 있다. 뇌졸중 후 마비말장애 디지털 치료기기 개발 사례를 중심으로 디지털 헬스 그리고 의료 인공지능에서의 설명가능한 인공지능의 역할과 중요성, 향후 나아가야 할 길에 대해 살펴본다.
As errors in medical AI can lead to critical consequences, the introduction of explainable AI has been actively pursued to address this issue. Focusing on the development of a digital therapeutic device for post-stroke dysarthria, this presentation examines the role and importance of explainable artificial intelligence in digital health and medical AI, and explores future directions for the field.
Invited Talk 5.
프론티어 AI 신뢰성 기술과 정책(Trustworthy use of the frontier AI: Technology and policy)
이재호 교수/CTO (서울시립대학교 인공지능학과/아리카)
의료 인공지능에서의 오류 발생이 치명적인 결과로 이어질 수 있는 분야인만큼 최근 이를 해결하기 위해 설명가능한 인공지능의 도입이 활발하게 이루어지고 있다. 뇌졸중 후 마비말장애 디지털 치료기기 개발 사례를 중심으로 디지털 헬스 그리고 의료 인공지능에서의 설명가능한 인공지능의 역할과 중요성, 향후 나아가야 할 길에 대해 살펴본다.
As errors in medical AI can lead to critical consequences, the introduction of explainable AI has been actively pursued to address this issue. Focusing on the development of a digital therapeutic device for post-stroke dysarthria, this presentation examines the role and importance of explainable artificial intelligence in digital health and medical AI, and explores future directions for the field.
Best Papers Presentation (09:40-10:10)
1.
물체 수준 시각 프롬프트를 활용한 효율적인 대형언어시각모델(CoLLaVO: Crayon Large Language and Vision mOdel)
이병관*, 박범찬, 김채원, 노용만 (KAIST)
Paper
Slides
2.
시간-주파수 기반 시계열 블랙박스 모델 설명(Time is Not Enough: Time-Frequency based Explanation for Time-Series Black-Box Models)
정현승*, 조수민, 권연수, 최윤재 (KAIST)
Paper
Slides
3.
대형언어모델에서 가이던스를 활용한 문서 암기 (Memorizing Documents With Guidance in Large Language Models)
박범진*, 최재식 (KAIST)
Paper
Slides
4.
인간 형상 그림 평가 및 그 외의 분야를 위한 부위별 기여도 평가 기반의 모델 설명 (PCEvE: Part Contribution Evaluation Based Model Explanation for Human Figure Drawing Assessment and Beyond)
이종서*, Geo Ahn*, 김성태*, 최진우 (경희대학교)
Paper
Slides
Oral Presentation 1 (10:55-11:50)
5.
음성에 기반한 마비말장애 진단과 설명이 가능한 시스템(A Speech-Based Dysarthria Diagnosis and Explanation System)
이정필*, 구명완 외 17명 (서강대학교), 김유영, 김진우 외 3명 (HAII), 송태진 (이대서울병원)
Paper
Slides
6.
커널 함수 및 태스크 속성 함수 기반 플러그앤 플레이 구조의 시각 모델 설명 가능 인공지능 방법(A Plug-and-Play Explainable AI for Multiple Task and Model of Visual Intelligence: An Approach Using Kernel and Attribute Functions)
조충상*, 김귀식, 박하영, 이영한 (KETI)
Paper
Slides
7.
확산 모델의 풀린 공간에서의 잠재벡터 조작을 통한 반사실 설명
(Counterfactual Explanation through Latent Manipulation in Disentangled Space of Diffusion Model)
나승협*, 이성환 (고려대학교 인공지능학과)
Paper
Slides
8.
뉴런 기반 디버깅을 위한 현실적인 이미지 공간에서의 클래스 조건부 특징 시각화(Class-conditional Feature Visualizations in Realistic Image Space for Neuron-based Debugging)
주영준*, 박지훈, 이성환 (고려대학교 인공지능학과)
Paper
Slides
9.
연관된 특성을 조작하여 결정경계를 조사하는 반사실 설명(Manipulating Related Features to Probe the Decision Boundary for Counterfactual Explanation)
김희동*, 이성환 (고려대 인공지능학과)
Paper
Slides
10.
설명가능한 인공지능을 위한 인공지능 시스템 적용 사례 분석(Use case Study of AI System for Explainability of AI System)
유인식*, 이재호 (서울시립대학교)
Paper
Slides
11.
시각적 반사실적 설명을 사용하여 흉부 X선 판별 모델의 가짜 상관관계 탐지(Using Visual Counterfactual Explanations to Detect Spurious Correlations in CXR Classification Models)
김원준*, 이상민, 예종철 (KAIST)
Paper
Slides
12.
실시간 임상 지원 시스템을 위한 깊은 합성곱 신경망을 이용한 설명가능 급성 신손상 예측 연구(Explainable AKI prediction with ResNet Toward Real-Time Clinical Decision Support)
성지현*, Enver menadjiev*, 한현경*, 최재식 (KAIST), 윤기애, 이진영, 김세중 (분당서울대학교병원)
Paper
Slides
13.
환자 중심의 의료 XAI 시스템 설계 및 검증(Advancing Healthcare AI: Patient-Centered Design and Validation of XAI Systems)
김유영*, 김민정, 김새별, 김진우 (주식회사 하이, 연세대학교 HCI Lab)
Paper
Slides
14.
미완성 프롬프트 탈옥(Incomplete Prompt Jailbreaking)
김연지, 이진실, 박범진, 정영주, 최재식 (KAIST)
Paper
Slides
15.
설명가능한 인공지능을 위한 사용자 중심의 설명 설계 가이드라인 개발: 체계적인 문헌 분석(Developing User-Centered Explanation Design Guidelines for Explainable AI: A Systematic Literature Review)
홍산*, 박경화, 박우진 (서울대학교)
Paper
Slides
Poster Presentation 1 (11:50-13:00)
16.
한국어 음성을 활용한 딥러닝 기반 성도 움직임 시각화(Deep Learning Based Visualization of Vocal Tract Movements Using Korean Speech)
Rodrigo Picinini Méxas, 음승호, 주윤지, 심윤섭, 박운상 (서강대학교)
Paper
17.
실측자료 정의를 통한 시계열 설명자 평가(Evaluating Time Series Explainers by Defining Ground Truth)
Enver Menadjiev, Artyom Stitsyuk, Tair Djanibekov, 최재식 (KAIST)
Paper
18.
설명 가능한 그래프 신경망의 최신 연구 동향(Recent Studies on Explainable Graph Neural Networks)
이재준*, 황지영 (KAIST)
Paper
19.
군사 도메인 분야 컴퓨터 비전 모델 개발을 위한 XAI를 활용한 domain generalization 사례(Domain Generalization Application Cases Utilized XAI Methods for Military Domain Vision Task Model)
이주현*, 정재혁, 김경수, 강현수 (코난테크놀로지), 김봉수 (육군교육사령부)
Paper
Oral Presentation 2 (14:40-15:40)
20.
설명가능한 인공지능(XAI)으로서 법률 LLM의 현재와 미래 : 법률 도메인의 특수성을 고려한 성능평가와 그 결과를 기초로(Present and Future of Legal LLM as Explainable AI(XAI) : Performance Evaluation Considering the Domain Specificity of the Legal Field)
이상엽*, 김소연, 강다영, 강필원, 이진원, 허남현, 홍대식 (서강대학교 법학전문대학원)
Paper
Slides
21.
안정된 확산모델의 속성 관련 특징지도 탐색(Analyzing the Attribute-relevant Featuremaps in Stable Diffusion Models)
권다희, 한지연 (KAIST), 최재식 (KAIST, INEEJI)
Paper
Slides
22.
대형언어모델에서 거절 메커니즘의 표현 해석(Representation Interpretation of Refusal Mechanism In LLMs)
박범진, 김연지, 이진실, 정영주, 최재식 (KAIST)
Paper
Slides
23.
조음교대운동을 이용한 마비말장애 중증도 자동 평가에서의 설명 가능 인공지능(Explainable Artificial Intelligence in Automated Dysarthria Severity Assessment using Oral Diadochokinesis)
김민서, 오준석, 김지환 (서강대학교)
Paper
Slides
24.
설명가능한 인공지능을 위한 인간 중심 평가 지표에 대한 체계적인 문헌 분석(A Systematic Literature Review of Human-Centered Evaluation Metrics for Explainable AI)
박경화, 홍산, 박우진 (서울대학교)
Paper
Slides
25.
인간 전문가를 모방한 도메인 전문 대규모 언어 모델 파이프라인(Domain Expert Large Language Model pipeline mimics a human domain expert)
한민수, 홍석영, 김수아, 조현민, 장현지, 김민주, 구명완 (서강대학교)
Paper
Slides
26.
대규모 언어모델에 기반한 질문 재작성을 활용한 검색 증강 생성 시스템(A RAG(Retrieval Augmented Generation) System using Query Rewritting Based on LLM(Large Language Model))
한민수, 홍석영, 정해인, 구명완 (서강대학교)
Paper
Slides
27.
Out-of-Distribution 탐지를 활용한 자연어 추론 학습데이터 생성을 위한 신규 규칙 발굴(Discovering Novel Rules for Generating Natural Language Inference Training Data Using Out-of-Distribution Detection)
한주영*, 황현선, 이창기 (강원대학교)
Paper
Slides
28.
이미지 분류기의 편향성 완화를 위한 텍스트 기반 마지막 레이어 재학습(TLDR: Text-based Last-Layer Retraining for Debiasing Image Classifiers)
박주현, 정석현, 문태섭 (서울대학교)
Paper
Slides
29.
ReDiff: 확산 분류기를 이용한 헤드리스 모델 리프로그래밍(ReDiff: Head-less Model Reprogramming with Diffusion Classifiers )
윤지석, 손정효, 정우택, 석흥일 (고려대학교)
Paper
Slides
30.
인과적 접근을 통한 확산 모델의 반사실적 이미지 이해(Understanding Counterfactual Image in Diffusion Models through a Causal Approach)
손정효*, 정우택, 송수정, 윤지석, 석흥일 (고려대학교)
Paper
Slides
31.
효과적인 대형 언어 모델의 미세조정(Effective methods for fine-tuning large-scale language models)
이형직*, 배경만 (전자통신연구원)
Paper
Slides
Poster Presentation 2 (15:40-16:10)
32.
그래프 경량에서 소거되는 변의 차수가 출력 그래프 품질에 미치는 영향(Influence of Degree in Edge Trimming for Graph Sparsification )
서현진*, 윤지훈*, 양은호 (KAIST)
Paper
33.
음성 기반 마비말장애 중증도 자동 평가 시스템(Speech-Based Dysarthria Severity Classification System)
김승진*, 김장연, 김민섭, 박형민 (서강대학교)
Paper
34.
XAI 기법을 이용한 확산 모델의 추론 메커니즘 분석(Analysis of Diffusion Model's Inference Mechanism Using XAI Techniques)
김도은*, 박인애, 변지은 (이화여대)
Paper
35.
112 긴급출동 의사결정 지원 시스템을 위한 설명 가능한 인공지능의 활용(Utilizing Explainable AI for Intelligent 112 Emergency Dispatch Decision Support Systems)
이정민*, 백명선 (한국전자통신연구원)
Paper
36.
입력변수 상호작용 기반의 이미지 분류기 설명(Explaining Image Classifiers Based on Interaction Analysis)
이명진*, 정성욱, 최재식 (KAIST)
Paper
37.
단락 검색을 위한 다중 관점 토큰 기반 늦은 상호작용 방법(Multi-View Tokens based Late Interaction for Passage Retrieval)
배용진*, 배경만 (ETRI)
Paper
38.
내재적 개념 병목 모델(Implicit Concept Bottleneck Model)
이찬우, 최재식 (KAIST)
Paper