협동과정 인공지능전공에서는 소속 학생들의 연구 및 교육 결과를 공유하고 교류의 자리를 마련하기 위하여 AI Lunch Seminar를 개최하고 있습니다. 

 

6월 23일(화) 12시에 진행되는 AI Lunch Seminar의 연사는 양기창(컴퓨터공학부 박사과정) 연구원입니다. 협동과정 인공지능전공 소속 학생 등 관심 있는 구성원의 많은 참여 바랍니다. 

 

◎ 일시: 2026년 6월 23일 화요일 12:00~13:00 

◎ 장소: 303동 661-6호(대회의실) 

◎ 발표자: 양기창(컴퓨터공학부 박사과정)

◎ 사전참가신청:   https://forms.gle/xhkSsVCUfeY4jwym6 6월 17일(수)까지 신청 부탁드립니다. 

※ 사전참가신청을 완료한 학생들에 한하여 점심 식사(서브웨이: 이탈리안 비엠티 OR 로스트 치킨)를 랜덤으로 제공할 예정입니다. 사전 신청 이후, 참석이 어려우실 경우에는 담당자에게 미리 연락 부탁드립니다.

 

◎ 발표 정보 

1. 주제

 I/O-Aware Systems for Real-Time Visual Intelligence

Real-time visual intelligence is increasingly deployed in resource-constrained and unstable environments, such as mobile robots, AR devices, and edge platforms. In these settings, the main obstacle is not only model computation, but also I/O: live video analytics suffers from unpredictable network loss and late packet arrivals, while edge VLM inference suffers from flash I/O latency caused by weight offloading. This seminar presents two systems that address these bottlenecks through I/O-aware model–system co-design. First, Logan enables loss-tolerant live video analytics by accepting packet loss and recovering corrupted video frames with codec-aware inpainting and fast-forward recovery, thereby improving latency SLO satisfaction with minimal accuracy loss. Second, VLM in a Flash introduces Neuron Chunking, an I/O-efficient sparsification method that selects contiguous neuron chunks by jointly considering activation importance and flash access cost. Together, these works show a common design principle: instead of treating I/O variability as an external constraint, real-time visual AI systems can expose and exploit model tolerance, data structure, and hardware behavior to achieve better accuracy–latency trade-offs.

 

3. Bio

Kichang Yang is a Ph.D. student in Computer Science and Engineering at Seoul National University. His research focuses on building efficient mobile and edge AI systems that can support real-time visual intelligence in resource-constrained and unstable environments.