Abu Dhabi Machine Learning

Register your interest – Abu Dhabi Machine Learning (ADML) Meetups

21st of May 2024 | Venue: ADGM Academy | Time: 3:00 PM to 5:00 PM

The Abu Dhabi Machine Learning (ADML) Meetups aim to gather a community with a passion for solving the technical challenges associated with using artificial intelligence techniques in real life settings. We aim to organize the Meetups monthly to explore the topic of machine learning at a detailed level.

The goal of these events is to keep growing the community’s knowledge of machine learning, its applications across several domains with a focus and relevancy to the financial sector, and its deployment best practices. Each meetup will end with networking time to discuss ideas and share contacts.

Event Agenda




3:00 PM to 3:40 PM

Learning to Learn Graph Topologies

In the rapidly evolving field of machine learning, the innovative paradigm of Learning to Learn (a.k.a. Learning to Optimise), has emerged as a notable approach, blending the adaptability of data-driven learning with the interpretability of rule-based optimisation. In this talk, Stacy will share insights from her NeurIPS publication titled "Learning to Learn Graph Topologies" (L2G), and its significant applications into the financial domain.

Stacy's research introduces an innovative methodological advancement in algorithmic unrolling, designed to refine the learning process for graph adjacency matrices. This approach is faster, more accurate, and offers increased flexibility compared to traditional methods, especially on rich topology representations and downstream task integration. The core of Stacy's presentation will be the L2G framework's adaptability to a broad range of downstream tasks, with a particular focus on a systematic trading strategy. In one of her follow-up paper, titled Learning to Learning Financial Network to Optimise Momentum Strategies, L2G framework has been applied to simultaneously learn graph topologies and optimise network momentum strategies for marco assets. The portfolio performance was gained from the flexible learning framework that can be trained with the negative Sharpe ratio, and accuracy capture the interconnections of these Marco assets.

Stacy Pu, Quantitative Researcher

Stacy Pu, Quantitative Researcher

3:40 PM to 4:20 PM

Harnessing Generative AI

As generative AI technologies rapidly advance, the business landscape is met with unprecedented opportunities and complex challenges. With over two thousand startups dedicated to generative AI, the market is bustling yet fraught with a gap in understanding between technology creators and business users.

This presentation delves into the crucial issues that stem from tech companies developing tools without a comprehensive grasp of their business applications and businesses struggling to recognise the tools’ value and assess associated risks.

Drawing analogies to oil exploration, deploying AI tools in business is discussed as a venture requiring significant capital investment and strategic navigation. The talk will address the need for tech companies to receive vital feedback and support from business users to refine and optimise AI applications.

Furthermore, it will emphasise the importance of businesses appointing in-house technology advocates and fostering both cross-industry and intra-industry cooperation to mitigate infrastructure risks.

The presentation will propose a pragmatic, step-by-step approach to AI tool implementation, focusing on practical daily applications, development of improvement requirements, and refinement processes to align with evolving business needs.

This session will provide attendees with insights into effectively integrating AI technologies into their operations, ensuring technological advancement and business growth.

Dmitriy Kaliada, Generative AI and Data Products Advisor

Dmitriy Kaliada, Generative AI and Data Products Advisor

4:20 PM to 5:00 PM

How Bad is Training on Synthetic Data? A Statistical Analysis of Language Model Collapse

The widespread integration of large language models like ChatGPT in late 2022 has led to a surge of synthetic content on the web, which poses a threat to the quality of future models due to a phenomenon known as model collapse. This issue arises when models are trained on a blend of human-generated and synthetic data, leading to a detrimental feedback loop. As synthetic outputs from one generation of models contaminate the training pool for the next, the quality and diversity of model outputs decline. This recursive train-generate loop, highlighted in studies by Shumailov et al. and others in 2023, results in models that produce increasingly narrow and repetitive outputs, forgetting the rich diversity of the original human-created data. The root causes of model collapse include statistical and functional approximation errors, which stem from the inherent limitations of models in capturing the full complexity of human language and the theoretical constraints of neural networks. This talk will present recent findings that underscore the critical need for carefully managing the ratio of synthetic to real data in training sets, in order to mitigate the risks of model collapse and preserve the linguistic diversity and effectiveness of future language models.

Mohamed El Amine Seddik, Lead Researcher, TII

Mohamed El Amine Seddik, Lead Researcher, TII

Registration Form