정보통신대학원 세미나를 소개합니다.
이번 학기는 온라인 or 오프라인으로 진행 되며, 학부생, 대학원생, 교직원 누구나 참여 가능합니다.
▶▶세미나2-정보통신대학원 세미나 2021년 2학기
Learning Architecutres and Algorithms for Collision Avoidance and Sensor Fusion
심명석 교수
(Shell Oil company, Machine Learning and AI (PhD in Electrical & Computer Engineering, Texas A&M University))
초록
The deep learning, which is a machine learning method based on artificial neural networks, enables cutting edge technologies such as autonomous driving, collision avoidance, activityrecognition, etc. While various machine learning algorithms are proposed for these applications,these algorithms suffer from limitations in terms of multi-modal sensor fusion, feature-level importance interpretability, and fusion architecture modeling. In this presentation, the aforementioned problems are addressed through the two proposed architectures: 1) a multiplicative reward modulated spike-timing-dependent plasticity (M-RM-STDP), which is a spiking neural network (SNN) based reinforcement learning; 2) an auxiliary model regulated gating architecture, called ARGate. In the spiking neural network based algorithm work, we review existing STDP based unsupervised learning methods, and additive reward modulated STDP(A-RM-STDP) based reinforcement learning in addition to classic Q-learning for mobile robot target tracking and collision avoidance as an application. Furthermore, a more biologically plausible feed-forward SNN architecture including coarser grained global rewards is examined.
Lastly, for the feature-importance analysis, fusion weights inconsistency, and noise resiliency, we discuss the proposed ARGate architecture with fusion weight regularization with auxiliary losses, and monotonic fusion target learning. The proposed gating network exceeds the baseline deep learning architectures with and without gating factorsin human activity recognition (HAR), driver identification, and KITTI datasets.
＊일시 9월 17일 (금), 13시
＊zoom https://handong.zoom.us/j/86244954271
＊회의 ID 862 4495 4271
＊문의 4단계 BK21 AI사업단 (260-3150)
