Foundational Math for Generative AI: Understanding LLMs and Transformers through Practical Applications

Course details

Unlock the mysteries behind the models powering today’s most advanced AI applications. In this course, instructor Axel Sirota takes you beyond just using large language models (LLMs) like BERT or GPT and highlights the mathematical foundations of generative AI. Explore the challenge of sentiment analysis with simple recurrent neural networks (RNNs) and progressively evolve your approach as you gain a deep understanding of attention mechanisms, transformers, and models. Through intuitive explanations and hands-on coding exercises, Axel outlines why attention revolutionized natural language processing, and how transformers reshaped the field by eliminating the need for RNNs altogether. Along the way, get tips on fine-tuning pretrained models, applying cutting-edge techniques like low-rank adaptation (LoRA), and leveraging your newly acquired skills to build smarter, more efficient models and innovate in the fast-evolving world of AI.

Learning objectives
Gain an intuitive understanding of how and why LLMs and transformers work.
Learn how attention mechanisms evolved to solve key problems in RNN-based models.
Develop a sentiment analysis model using TensorFlow, Keras, and Hugging Face’s DistilBERT.
Enhance models progressively with mathematical insights applied in code, from word embeddings to attention and transformer layers.
Use visualizations to grasp how attention and optimization work.

Concepts

Introduction

Intro to foundational math for generative AI
Getting the most out of this course
Version check

Introduction to Math for GenAI and Attention Basics

Why LLMs and attention matter
RNNs and the context bottleneck problem
Demo - Building a simple RNN model for sentiment analysis
Introduction to attention - Bahdanau s solution
Demo - Adding attention to an RNN model
Solution - Implement Bahdanau's attention

Transformers - Removing RNNs for More Efficient Models

From RNNs to transformers
Understanding self-attention in transformers
Multi-head attention and positional encoding
Building a transformer model for sentiment analysis
Solution - Build a two-layer transformer encoder

Deep Dive into LLMs and Model Fine-Tuning

The three types of LLMs
Special decoder-only models
Explaining encoder-only models like BERT
Fine-tuning DistilBERT for sentiment analysis
Attention masks in transformers
Solution - Detect irony and climate stance in TweetEval

Conclusion

Course summary and next steps