Tutorial 1
Graph-Based Indexes for Approximate Nearest Neighbour Search
Gold Coast Time: Tue 16 Dec 2024 15:30 AEST (UTC+10)
Tokyo Time: Tue 16 Dec 2024 14:30 JST (UTC+9)
▶  Abstract
Nearest Neighbour Search is a critical technique for queries involving vectors, however due to the Curse of Dimensionality modern solutions scale incredibly poorly for high dimensional vectors. As a result, the problem of Approximate Nearest Neighbourhood Search has become the most practical solution for high dimensional datasets, sacrificing the need for perfect accuracy in exchange for reasonable runtime. A common method to improve the efficiency of Approximate Nearest Neighbour Search is the adoption of an index, with one of the most popular options being a Graph-Based Index. These solutions are attractive as they can effectively provide a comparable level of usability regardless. In this tutorial we will explore various graph-based indexes proposed over time which declare to be a usable option. To do so we will examine the foundational ideas behind their models as well as discuss the shortcomings behind their designs. We will also raise questions regarding the future direction of graph-based indexes and their practicality for vector databases.
▶  Speaker
Zhou Alexander
Hong Kong University of Science and Technology
Zhou Alexander is a Research Assistant Professor at the Hong Kong Polytechnic University. He received his PhD in Computer Science at the Hong Kong University of Science and Technology and prior to that received a Bachelor of Engineering and Bachelor of Science at the University of Queensland. He specialises in algorithms for graph databases with a particular interest in social science driven design and has published in multiple top venues such as ICDE, VLDB, SIGKDD, VLDBJ, TKDE and more. He also serves as committee member and reviewer for conferences and journals such as ICDE, WWW, DASFAA, TODS, ADMA and ADC.
Tutorial 2
Robustness Verification of Neural Networks using WraLU
Gold Coast Time: Tue 16 Dec 2024 16:30 AEST (UTC+10)
Tokyo Time: Tue 16 Dec 2024 15:30 JST (UTC+9)
▶  Abstract
The increasing deployment of deep neural networks (DNNs) in safety-critical domains has made the formal verification of their robustness an argent task. A common approach to DNN verification involves discomposing the network into multiple layers of neurons, where linear and non-linear operations (activation functions) stack to form the network’s overall function. The key challenge lies in constructing linear constraints that effectively bound those non-linear activation functions – a technique known as function hull over-approximation. In this tutorial, we introduce WraLU (Wrapping ReLU), our novel approach for computing the convex hull of the ReLU function. WraLU enables the construction of multi-neuron constraints for verifying ReLU-based DNNs. It significantly reduces computation time and constraint number to improving efficiency in high-dimensional settings.
▶  Speaker
Guangdong Bai is an Associate Professor in the School of Electrical Engineering and Computer Science at the University of Queensland, Australia. His research spans responsible machine learning, security, and privacy. He is an Associate Editor of IEEE Transactions on Dependable and Secure Computing.
Zhongkui Ma is currently a PhD student in the School of Electrical Engineering and Computer Science at the University of Queensland, Australia. His research interest includes formal verification and machine learning.
Tutorial 3
LLM Acceleration through Speculative Decoding
Gold Coast Time: Tue 17 Dec 2024 15:30 AEST (UTC+10)
Tokyo Time: Tue 17 Dec 2024 14:30 JST (UTC+9)
▶  Abstract
Large language models (LLMs) have become increasingly powerful and are now adopted for a wide range of applications. Many LLM applications require real-time responses. However, LLMs are typically based on the autoregressive transformer architecture, which generates output tokens step by step and thus suffers from high latency, particularly with larger model sizes. To reduce LLM serving latency, speculative decoding has been proposed as a viable solution, where a small draft model produces speculated output sequences that are verified by the large target model in parallel to achieve a faster decoding speed. In this tutorial, I will first introduce existing work on speculative decoding. I will then present our recently proposed speculative decoding method called GliDe with a CaPE that introduces two low-hassle modifications to vanilla speculative decoding to improve the speedup.
▶  Speaker
Jing Jiang
Australian National University
Jing Jiang is a Professor in the School of Computing at the Australian National University. Previously she was a Professor and Director of the AI & Data Science Cluster in the School of Computing and Information Systems at the Singapore Management University. Her research interests include natural language processing, text mining, and machine learning. She has received two test-of-time awards for her work on social media analysis, and she was named Singapore’s 100 Women in Tech in 2021. She holds a PhD degree in Computer Science from the University of Illinois Urbana-Champaign.
Tutorial 4
Dynamic Routing in Road Networks
Gold Coast Time: Tue 17 Dec 2024 16:30 AEST (UTC+10)
Tokyo Time: Tue 17 Dec 2024 15:30 JST (UTC+9)
▶  Abstract
Route planning is the fundamental operation for the navigation services and the building block of various location-based applications. In this tutorial, a batch of dynamic routing algorithms will be systematically introduced in terms of the index construction, query processing, index update and their applicable scope. In addition, multiple related future research directions will be briefly discussed.
▶  Speaker
Mengxuan Zhang
Australian National University
Mengxuan Zhang is a Lecturer from the Australian National University (ANU). Before joining ANU, she worked as a postdoc at Iowa State University and a Research Associate at the Hong Kong University of Science and Technology. Mengxuan obtained her PhD from the University of Queensland in 2021, under the supervision of Prof. Xiaofang Zhou and Dr. Lei Li. Her research interests include Database, High-performance Query Processing and Route Planning.