In this TechTalk, we will explore the challenges and best practices for developing ethical and responsible Large Language Models (LLMs). The talk will delve into the importance of transparency and explainability, showcasing methods such as attention visualization and model distillation that provide critical insights into model behavior. We will address the control of generated content through techniques including reinforcement learning from human feedback (RLHF), token penalization, external moderation systems, and prompt engineering. Finally, we will tackle the issue of bias mitigation, emphasizing the need for transparency in the pre-training data used for these billion-parameter models, diverse and representative data, as well as pre-, in-, and post-processing techniques to ensure fairness in LLMs.
These considerations have taken on heightened significance in light of the recently proposed EU AI Act, the first major regulatory framework for AI. Although the Act originates in Europe, its reach is globally impacting organizations that interact with European entities or serve European customers. In addition, it might set a precedent that could inspire similar legislation in other regions, thereby broadening its implications.
Moreover, understanding these critical aspects of responsible AI not only aids in the creation of LLMs that are ethical and safe for real-world applications but also accelerates the development process for downstream tasks. Since these techniques can be used as an aid for developers to swiftly identify and correct issues, improving model performance and reliability.
Nicole Königstein is a distinguished Data Scientist and Quantitative Researcher, currently working as Chief Data Scientist and Head of AI & Quantitative Research at Wyden Capital. Alongside her role in this organization, she serves as an AI consultant across diverse industries, leading workshops and guiding companies from the conceptual stages of AI implementation through to final deployment. As a guest lecturer, Nicole shares her expertise in Python, machine learning, and deep learning at various universities. She is a regular speaker at renowned AI and Data Science conferences, where she conducts workshops and educational sessions. In addition, she is an influential voice in the data science community, regularly reviewing books in her field and offering her insights and critiques. Nicole is also the author of the well-received online course, Math for Machine Learning, published by Manning Publications and the author of the (forthcoming) book Transformers in Action published by Manning Publications.
Miquel Noguer i Alonso is a seasoned financial professional and academic with over 30 years of experience in the industry. He is the Founder of the Artificial Intelligence Finance Institute and Head of Development at Global AI. His career includes roles such as Executive Director at UBS AG and CIO for Andbank. He has served on the European Investment Committee UBS for a decade. He is on the advisory board of FDP Institute and CFA NY. Miquel is also an academic, teaching AI, Big Data, and Fintech at institutions like NYU Courant Institute, Columbia University, and ESADE. He pioneered the first Fintech and Big Data course at the London Business School in 2017. He is the author of 10 papers on Artificial Intelligence. He holds an MBA and a Degree in business administration from ESADE, a PhD in quantitative finance from UNED, and other prestigious certifications. His research interests span asset allocation, machine learning, algorithmic trading, and Fintech.
