An IEEE CS Türkiye Chapter Lecture Series by Prof. John Abela
An IEEE Computer Society Türkiye Chapter LLMS Lecture Series in collaboration with Özyeğin University and University of Malta
We are pleased to announce the LLMs Talk Series organized by the IEEE Computer Society Türkiye Chapter, in collaboration with Özyeğin University and the University of Malta.
Speaker: Prof. John Abela
Associate Professor, Faculty of ICT, University of Malta
Location: Özyeğin University, Çekmeköy Campus, Istanbul
Dates: September 29 – October 3, 2025
🎯 About the Lecture Series
- This seminar series is designed for senior, graduate and postgraduate students, postdocs, and faculty who are interested in gaining a deeper understanding of how Large Language Models (LLMs) work.
- A solid background in Artificial Intelligence (AI) and Neural Networks is required.
- The series will provide both theoretical insights and practical foundations, with a strong focus on the OpenAI GPT-3 model. Participants will explore the technical mechanisms that enable large-scale neural architectures to achieve state-of-the-art performance in Natural Language Processing.
📚 Topics to be Covered
1) Introduction to NLP Foundations – from symbolic systems to modern connectionist models.
2) Text Representation Journey – one-hot encodings, count-based models, and distributed embeddings.
3) Deep Dive into Transformers – explore the architecture powering ChatGPT and GPT-3.
4) Inside GPT-3 – dissect the Transformer paradigm and its groundbreaking mechanisms.
5) Training at Scale – how massive datasets and attention mechanisms capture complex language patterns.
6) Math Behind the Models – why vector arithmetic and tensor algebra are critical for understanding LLMs
⏰ Schedule
- 29 September 2025
- 09:40 – 11:30
- 14:40 – 16:30
- 30 September 2025
- 11:40 – 12:30
- 13:40 – 14:30
- 1 October 2025
- 14:40 – 16:30
- 2 October 2025
- 09:40 – 11:30
- 3 October 2025
- 09:40 – 11:30
👨🏫 About the Speaker
Prof. John Abela is an Associate Professor in the Faculty of ICT at the University of Malta. He holds a BSc in Mathematics and Computing (Malta), an MSc in Computer Science (University of New Brunswick, Canada), and a PhD in Theoretical Machine Learning (UNB).
He co-founded a successful ICT company in 2003, which he exited in 2019. His research interests include Artificial Intelligence, Machine Learning, Deep Learning (including Transformer models), Image Processing, Natural Language Processing, and Large Language Models.
📝 Registration
Participation is free of charge, but registration is required.
For registration please CLICK HERE.
Abstract – How Large Language Models (LLMs) Work
The rapidly evolving landscape of Natural Language Processing (NLP) has been transformed by the advent of neural network architectures and, in particular, the development of large language models (LLMs) based on Transformer technology. This lecture series offers an in-depth exploration of the theoretical and practical underpinnings that make modern models such as ChatGPT and other advanced LLMs possible.
The course begins with an introduction to the core principles of NLP and the evolution from symbolic systems to connectionist models. Traditional vector space methods, which treat data as numerical representations within high-dimensional spaces, serve as the foundation for many modern AI algorithms. We discuss the rationale behind representing text as vectors and outline the advantages that continuous representations offer over classical symbolic approaches. This initial section also revisits linear algebra and the fundamental machine learning concepts that are essential for understanding how vector space models function in NLP contexts.
Particular emphasis is placed on word embeddings—a breakthrough that redefined how semantic and syntactic relationships are captured numerically. We will trace the evolution of textual representations from one-hot encodings and count-based models to distributed representations. In particular, we focus on the Word2Vec model, its mechanisms (such as the Continuous Bag-of-Words and Skip-Gram training methods), and its advantages in learning semantic relationships from vast text corpora. This discussion extends to alternative embedding methods, including GloVe, FastText, and contextual embedding models such as BERT, which incorporate dynamic word representations based on context. The objectives include understanding both the historical context and the practical challenges of embedding words as numeric values.
In the last two sessions of the series, we deep dive into the architecture of Transformer models — the driving force behind LLMs such as ChatGPT (version 3.5), and emerging successors. The sessions will dissect the Transformer paradigm – in particular the GPT-3 Transformer model from OpenAI. We explain, in quite some detail, positional encodings, batch creation, the transformer block including the multi-head self-attention mechanism, and the linear projection layer. Participants will study the technical nuances of model training, including the role of multi-head attention, feed-forward layers, and the significance of residual connections and normalization. Detailed analysis of GPT-3’s architecture, with its 175 billion parameters and the implications of scaling, will be provided. Moreover, emerging properties such as few-shot learning and the interplay between scale and model capabilities will be critically examined, with an outlook on innovations such as Mixture of Experts (MoE) and potential future developments.
The curriculum is designed to bridge the gap between the theoretical and algorithmic foundations of the self-attention transformer architecture. Students will gain insight into how large-scale models are trained on massive datasets, how attention mechanisms empower these models to capture intricate language patterns, and why a comprehensive understanding of underlying mathematics—from vector arithmetic to tensor algebra—is critical in the development and interpretation of modern LLMs
English
0 Comments