AIAI Seminar -17 June-Talk by Patrick Kage, Magdalena Proszewska, Yifei Xie

Speaker: Patrick Kage

Title: Representation Learning for Geospatial Analysis

Abstract: Geospatial imagery analysis is an increasingly important tool in understanding our rapidly changing world, with applications ranging from disaster response and weather forecasting to urban planning and resource management. A major challenge in applying traditional machine vision techniques to this problem space is the relative scarcity of labeled data and the inherent differences between natural imagery and satellite imagery. In this talk, I will outline the state of the art of unsupervised representation learning as applied to geospatial data and present my current research towards adapting vision transformers for multiscale and multimodal representation learning.

Speaker: Magdalena Proszewska

Title: Graph Kernel Convolutions for Interpretable Classification

Abstract: State-of-the-art Graph Neural Networks (GNNs) have demonstrated remarkable performance across diverse domains, hence the growing demand for more interpretable GNN techniques. While current research predominantly centers on post hoc perturbation techniques, recent studies propose use of Graph Kernel Convolutions (GKConv) to increase GNNs interpretability intrinsically. These models employ trainable graph filters for extracting hidden features. After training, they are investigated to uncover patterns present in the data that allow to make predictions. However, interpretability of such models is limited since they heavily rely on multilayer perceptrons (MLPs). In our paper, we show that it is possible to  build a model that solely relies on graph kernels and a simple linear layer. Additionally, we integrate contrastive loss to encourage the learning of a more descriptive set of graph filters, where each filter can be associated with one of the classes. In consequence, its decision-making process described through found graph filters and said linear layer is more interpretable. As a proof of concept, we propose a shallow GKConv Interpretable Classifier, which is able to achieve state-of-the-art results while exhibiting better interpretability.

Speaker: Yifei Xie

Title: A view change optimization model for leader-based parallel Byzantine consensus algorithm.

Abstract: The development of blockchain has driven extensive research on Byzantine Fault-Tolerant (BFT) consensus algorithms. Among them, leader-based parallel BFT algorithms have become particularly favoured in practical applications due to their superior performance. These algorithms rely on the view-change protocol to ensure system operations by detecting a faulty leader replica and initiating the selection of a new leader through the consensus procedure. While prior research focused on optimizing the replication protocol to enhance system performance, the importance of the view-change protocol has often been underestimated. Recently, however, the view-change protocol has emerged as a critical factor in system performance, especially as view changes occur more frequently than previously assumed. Traditional view-change protocols assume that all replicas blindly follow a predefined schedule and rotate leadership, leading to weak robustness and inefficiency. To address this issue, our work proposes a view-change optimization model that addresses the weaknesses inherent in passive view-change protocols, thereby improving the performance and resilience of leader-based parallel BFT algorithms.