AI systems have recently achieved state-of-the-art performance and are now widely used across many aspects of our daily lives. However, these models often struggle when encountering data that differs significantly from what they were trained on—a phenomenon known as distribution shift. For example, a self-driving car trained in urban environments may perform poorly when deployed in rural settings. In this talk, I will cover (1) how AI systems are trained and make predictions, (2) different types of distribution shifts and the challenges they pose, and (3) recent approaches developed to address these issues.

In smart mobility, vast networks of geographically distributed sensors generate high-frequency spatio-temporal data that must be processed in real time to avoid disruptions. Centralized approaches struggle to scale with growing sensor networks and are prone to reliability issues. To address these challenges, we adapt semi-decentralized training techniques for Spatio-Temporal Graph Neural Networks (ST-GNNs). Sensors are grouped by proximity into cloudlets, each managing a subgraph, exchanging node features, and sharing model updates to ensure consistency, eliminating reliance on a central aggregator. We evaluate centralized, traditional FL, server-free FL, and Gossip Learning setups on METR-LA and PeMS-BAY datasets for short-, mid-, and long-term predictions. Results show  that semi-decentralized setups are comparable to centralized methods in performance while improving scalability and reliability, with attention to communication and computational costs often overlooked in existing literature. 

This tutorial provides a step-by-step guide to time series forecasting. The first part is aimed at beginners, demystifies the core components of time series data, such as trend, seasonality, and noise, and guides through the essential steps of preparing data for predictive modeling.  
The second part analyses the most common forecasting techniques, from classical statistical methods to modern machine learning to deep learning approaches. Each method is illustrated with examples to help understand how and when they can be used effectively.

This presentation introduces a distributed architecture for real-time multi-object tracking across multiple camera streams, optimized for deployment on resource-constrained Jetson devices. We present a fully asynchronous, multi-threaded pipeline that balances computational workloads across capture, preprocessing, TensorRT-based YOLOv8 inference, feature embedding extraction, and DeepSORT tracking. Emphasis is placed on low-level system optimization, including parallelism, memory management, and pipeline parallelism to maximize performance on heterogeneous edge hardware. 
To support scalable cross-camera Re-Identification (Re-ID), we decouple identity resolution from edge inference by transmitting compact appearance embeddings to a central Kafka broker. These embeddings are indexed using a high-performance vector database (e.g., FAISS), enabling efficient global identity assignment while minimizing network bandwidth. We conclude with performance evaluations, design trade-offs, and insights into building scalable edge-to-cloud tracking systems under strict hardware constraints. 

This tutorial introduces participants to the principles and practices of efficient deep learning, with a focus on core model compression techniques such as pruning, quantization, and the design of efficient neural architectures. Through a hands-on walkthrough, participants will learn how to apply fine-grained pruning and perform model quantization. The tutorial is designed to provide a solid introduction and practical starting point for developing efficient, low-footprint AI systems suitable for real-world use.