Optimizing Memory Efficiency in SSL Video Models: A Scalable Reversible Paradigm Approach; Championing Digital Privacy: The Role of Advocacy Groups in the AI Era
Alghannam, Abdulmohsen, School of Engineering and Applied Science, University of Virginia
Vrugtman, Rosanne, EN-Comp Science Dept, University of Virginia
Morrison, Briana, EN-Comp Science Dept, University of Virginia
Norton, Peter, EN-Engineering and Society, University of Virginia
Stafford, William, EN-Engineering and Society, University of Virginia
In this thesis, I explore two significant areas within the realm of computer science and technology policy. The technical aspect of my thesis addresses the optimization of memory efficiency in Self-Supervised Learning (SSL) video models through a scalable reversible paradigm approach. Meanwhile, my Science, Technology, and Society (STS) paper delves into the critical role of advocacy groups in championing digital privacy in the age of artificial intelligence. Although these projects may initially appear distinct, they are intricately linked by their implications for ethical technology use and the overarching goal of responsible innovation.
The primary focus of the technical component of my thesis is on optimizing memory efficiency in Self-Supervised Learning (SSL) video models through innovative architectural designs. In collaboration with the Image and Video Understanding Lab at King Abdullah University for Science and Technology, we have developed the Dynamic Reversible Dual-Residual Networks (Dr2Net). This architecture combines the benefits of reversible networks with traditional SSL paradigms to address the significant challenge of high memory consumption during the model training phase. Reversible models offer the advantage of discarding non-essential data during forward passes, thereby reducing the memory footprint. The Dr2Net leverages this by integrating these reversible mechanisms with pre-trained SSL models, enabling them to handle larger datasets and more complex video processing tasks with less memory. The aim is to produce a scalable solution that maintains or improves accuracy, facilitates larger scale implementations, and supports the growing demand for advanced video analysis tools in fields such as surveillance, media, and autonomous vehicles. This approach not only enhances the technical capabilities of SSL video models but also contributes to the broader discussion on sustainable AI development in resource-constrained environments.
The STS paper explores the critical role of advocacy groups in the era of artificial intelligence, particularly focusing on their efforts to champion digital privacy rights. As AI technologies become increasingly integrated into everyday life, they bring with them substantial risks related to data privacy and surveillance. This paper assesses how organizations like the Electronic Frontier Foundation (EFF) and the Electronic Privacy Information Center (EPIC) engage in policy debates, litigation, and public discourse to safeguard individual privacy against potential overreach facilitated by AI technologies. It provides an analysis of major privacy issues raised by the deployment of advanced AI systems, including those related to data collection, analysis, and storage, which are particularly pertinent to SSL video models like those discussed in the technical portion of the thesis. By drawing on case studies and current regulatory challenges, the paper highlights effective strategies employed by these groups to influence policy and promote a balanced approach to AI development. This involves ensuring that AI respects human rights and adheres to ethical standards, thereby preventing its use in ways that could undermine personal autonomy and democratic freedoms. The study underscores the necessity for ongoing vigilance and advocacy to shape the development of AI systems in a manner that aligns with societal values and legal norms, reflecting a broader imperative to manage technological advancement responsibly.
These two projects converge on the ethical deployment and development of AI technologies. Both segments acknowledge the power and potential of AI to transform societies and industries but also caution against the unchecked use of such technologies. The technical project contributes to this dialogue by proposing a model that not only enhances performance but also reduces the demand on computational resources, potentially decreasing the barriers to privacy preservation in data-intensive tasks. Similarly, the STS paper stresses the need for robust oversight and ethical frameworks to govern AI development, highlighting how advocacy can influence the trajectory of technology towards more equitable and privacy-conscious practices.
While the technical project and the STS paper may focus on different aspects of AI and technology, they converge on the critical theme of ethical responsibility in technological advancement. By examining both the technological and societal implications of AI, this thesis aims to contribute to a broader understanding of how technology can be developed and governed to foster not only innovation but also a just and privacy-respecting society. Through these investigations, we see that effective advocacy and innovative technical solutions are both crucial for shaping the future of responsible AI usage.
BS (Bachelor of Science)
Digital Privacy, Machine Learning, Artificial Intelligence, Transformers
School of Engineering and Applied Science
Bachelor of Science in Computer Science
Technical Advisor: Briana Morrison, Rosanne Vrugtman
STS Advisor: William Stafford
Technical Team Members: Chen Zhao, Shuming Liu
English
All rights reserved (no additional license for public reuse)
2024/05/09