AI, ML and Friends is a weekly seminar series within the School of Computing on Artificial Intelligence, Machine Learning, and related topics. We are open to attendees and presenters external to the school. Please sign up to the mailing list to receive weekly announcements including zoom details, and email the seminar organiser to schedule a talk.

Upcoming Seminars #

01 August 2024, 11:00 #

Effective and Scalable Representation Learning for Graphs: From Theory to Applications #

Speaker: Hao (Allen) Zhu

Abstract: This presentation demonstrates critical challenges in graph-based machine learning, focusing on oversmoothing, effectiveness, and scalability. We introduce Simple Spectral Graph Convolution (S²GC), a variant of Graph Convolutional Networks with improved computational efficiency and performance. We then extend classical Laplacian Eigenmaps to develop Contrastive Laplacian Eigenmaps (COLES) and Generalized Laplacian EigeNmaps (GLEN) for unsupervised graph representation learning, offering scalable solutions for large-scale graph embedding with performance guarantees. Finally, we explore applications in few-shot learning and dynamic graph construction, introducing Unsupervised Discriminant Subspace Learning (EASE) and Prototype-based Label Propagation (ProtoLP). These innovations significantly advance the field of graph representation learning, offering improved performance and efficiency across various tasks including node classification, clustering, and transductive inference.

Bio: Hao (Allen) Zhu is a PhD student in his final year at the School of Computing at ANU. Her research involves graph machine learning on theories and applications. During his doctoral studies, Allen has authored 12 research papers that have been published in ML, DM, CV conferences.

Where: Building 145, room 1.33

08 August 2024, 11:00 #

General Keypoint Detection: Few-shot, Zero-shot and Beyond #

Speaker: Changsheng Lu

Abstract: This talk presents the advancements in computer vision and machine learning through the lens of keypoint detection, especially relating to few-shot learning, uncertainty learning, vision transformers, and multi-modal foundation models. Existing machine learning models rely heavily on extensive human-labelled data, leading to the problem of “no intelligence without human annotations”. To ease this issue and break the limitation of keypoint types to be detected, we introduce a versatile Few-shot Keypoint Detection (FSKD) with uncertainty learning, which can not only detect a varying number of keypoints of different kinds but also provides uncertainty estimation. Secondly, building robust keypoint representations is crucial to the success of FSKD. To achieve this, we propose a novel visual prior guided vision transformer and explore i) transductive extension of FSKD and ii) FSKD with masking and alignment (MAA). Thirdly, despite the versatility of existing FSKD models, they suffer from the scalability issue w.r.t. the number of keypoints and the large domain shift of keypoints between seen and unseen species. To address these issues, we propose a lightweight FSKD model capable of evaluating a large number of keypoints (e.g., 1000 keypoints) and improve our model with mean feature based contrastive learning to bridge the domain shift. Fourthly, we observe that our FSKD aligns with the popular “prompt” based models used across various vision and language tasks, if treating the support image and keypoints as “visual prompt”. Thus, we expand the prompt diversity from three aspects: modality, semantics (seen v.s. unseen), and language, to enable a more generalized zero- and few-shot keypoint detection (Z-FSKD). Finally, we demonstrate the FSKD beyond supervised setting and discuss the future directions for general keypoint detection.

Bio: Changsheng Lu is a PhD student in his final year at the School of Computing, ANU. He obtained his M.S. degree in control science and engineering from Shanghai Jiao Tong University, Shanghai, China, in 2020, and his B.S. degree in Automation from Southeast University, Nanjing, China, in 2017. He has wide research interests in computer vision, machine learning, and robotics. He is recently working on the general keypoint detection, multimodal foundation model, zero- and few-shot learning, and transfer learning. Particularly, he is interested in the theories and algorithms that can empower robot to see, think and conduct more like a human. Previously, he was awarded the national scholarship and the outstanding graduate student of SEU and SJTU. He has served as the reviewer for journals including T-PAMI, IJCV, T-IP, PR, IEEE RA-L, etc., and conferences including CVPR, ICCV, ECCV, AAAI, NeurIPS, ICML, ICLR, etc.