Presenter:
Engin Zeydan
- Senior Researcher, Centre Tecnològic de Telecomunicacions de Catalunya (CTTC)
Abdullah Aydeger
- Department of Electrical Engineering and Computer Science, Florida Institute of Technology

Abstract:

The tutorial proposal focuses on the convergence of quantum threats in the domain of 6G networks. It aims to provide an in-depth study of this convergence, starting with background information on quantum attacks, post-quantum cryptography, and quantum key distribution. It will then explore its execution to 6G networks and their quantum-based threats. The tutorial will include a step-by-step demonstration of two of the demos to illustrate the practical implementation of these concepts. The tutorial is designed for participants with no prerequisite knowledge and aims to introduce them to the application of post-quantum cryptography and quantum key distribution to protect the 6G networks. As this topic is gaining significance and relevance in the telecommunications industry, the tutorial offers attendees the opportunity to learn about cutting-edge security issues for 6G networks and their specific applications from the cybersecurity perspective.

Presenter:
Shiwei Liu
- University of Oxford
Arijit Ukil
- TCS Research

Abstract:
Large language models (LLMs) have become pivotal in modern deep learning, making it essential to understand the underlying patterns, particularly as these models scale exponentially. With parameter counts skyrocketing from billions to trillions in recent years, the associated computational costs and energy consumption for training and fine-tuning have escalated dramatically. This rapid expansion has sparked a growing interest in techniques that can achieve more compact model paradigms without sacrificing performance. Sparsitybased model compression has emerged as one of the most promising approaches for optimizing model efficiency, yet its application to LLMs has lagged behind other compression strategies. To address this gap, we revisit the existing landscape of model compression, focusing specifically on sparse neural networks, and offer a clear categorization of the methodologies employed in this space. We then explore recent advancements in sparsity for LLM compression and the caveat of sparsity in LLMs. The tutorial will conclude with identifying key challenges and opportunities in this field. Ultimately, this tutorial provides a comprehensive roadmap tracing the evolution of model compression and sparsity from traditional models to large-scale LLMs, emphasizing its critical importance in the ongoing development of efficient and scalable AI systems.

Presenter:
JJ Merelo-Guerv'os
- Department of Computer Engineering, Automatics and Robotics, University of Granada, Spain

Abstract:
Green computing is a general term that describes a host of techniques that try to minimize the carbon footprint of software applications. As such, it is not a single body of knowledge, but a series of best practices that help reduce energy consumption relying on the features of any of the different layers that are exercised by software applications. This represents a challenge at the time of designing a comprehensive syllabus that would help students develop the series of skills needed to identify energy bottlenecks and eliminate them. In this poster we will describe the different concepts involved, and how they will be delivered to guarantee the achievement of learning objectives. Green computing [2] deals, in general, with reducing the environmental impact of the creation and use of computing resources. From the software perspective, it proposes maximizing the amount of work done for every unit of energy spent. But to achieve that, how energy is spent across all the different computing layers need to be assessed and understood. This is why getting the student to achieve a certain amount of understanding of the different process involved, methodologies needed to carry out that assessment, and eventually design your code from the ground up or refactoring it to make it greener is a challenge.

Presenter:
Mike Mannion
- Department of Computing, Glasgow Caledonian University, Glasgow, UK
Hermann Kaind
- Institute of Computer Technology, TU Wien, Vienna, Austria

Abstract:
The volume, variety and velocity of products in software-intensive systems product lines is increasing. One challenge is to understand the range of similarity between products. Reasons for product comparison include (i) to decide whether to build a new product or not (ii) to evaluate how products of the same type differ for strategic positioning or branding reasons (iii) to gauge if a product line needs to be reorganized (iv) to assess if a product falls within the national legislative and regulatory boundaries. We will discuss two different approaches to address this challenge. One is grounded in feature modelling, the other in case-based reasoning. We will also describe a specific product comparison approach using similarity matching, in which a product configured from a product line feature model is represented as a weighted binary string, the overall similarity between products is compared using a binary string metric, and the significance of individual feature combinations for product similarity can be explored by modifying the weights. We will illustrate our ideas with a mobile phone example, and discuss the benefits and limitations of this approach.

Title: Advancements in Edge AI for ASR

Abstract:
We present a novel Automatic Speech Recognition (ASR) model developed to operate efficiently on low-powered edge compute devices, such as those found in car infotainment systems. Our model is based on the conformer architecture and has been fine-tuned using a diverse training dataset.
The key innovation of our approach lies in the finetuning and the optimization techniques employed to enable the model to run on resource-constrained edge devices. Despite these optimizations, our model achieves evaluation results that surpass existing open-source and commercial ASR models.
We believe that our work represents a significant advancement in the field of edge AI and has the potential to revolutionize the way speech recognition is performed in a variety of applications, including in-car voice control.

Title: The Pursuit of Privacy in the AI Age

Abstract:
We generate and share vast amounts of data that reveal a detailed portrait of our lives, exposing our identity, behaviors, and preferences. To enable individuals to exercise greater control over their personal information, I will present novel approaches to identify and protect sensitive attributes within data. I will present feature representations that effectively disentangle sensitive information from non-sensitive attributes. Furthermore, I will present perturbation techniques designed to obfuscate sensitive attributes while preserving or potentially enhancing the overall quality of the data, thereby safeguarding sensitive information from unwanted inference.