AI MASTERY HANDBOOK: FROM DEEP LEARNING FUNDAMENTALS TO CONQUERING LLMs

Synopsis

The AI MASTERY HANDBOOK is meticulously crafted for students at R.I.S.E. CENTER, providing a comprehensive roadmap from foundational AI concepts to advanced Large Language Models (LLMs). This textbook bridges the gap between theoretical knowledge and practical application, empowering readers to become combat-ready AI engineers. It systematically guides you through the essential pillars of artificial intelligence, starting with the core principles of deep learning. You'll gain a solid understanding of neural networks, data preprocessing, model training, and evaluation techniques, building a robust foundation for more complex AI systems. As you progress, the handbook delves into the intricacies of modern AI, with a dedicated focus on conquering LLMs. Explore the architecture, training methodologies, and deployment strategies of state-of-the-art language models. Case studies and real-world examples illustrate how these powerful tools are transforming industries. Whether you're starting from zero or looking to deepen your expertise, this handbook offers the clarity, structure, and actionable insights needed to master AI and excel in this rapidly evolving field. STYLE GUIDE: Theory (40%): Explain principles deeply but accessibly; prioritize visual thinking (describing diagrams, data flows) over walls of text. Mathematics (20%): Do not shy away from formulas. Explain the physical/geometric meaning of the formulas (e.g., Gradient Descent is like stepping down a mountain while blindfolded). Use standard LaTeX for equations. Code (40%): Focus on practice. Sample code must be in PyTorch, structured as Jupyter Notebooks (cell by cell), with line-by-line instructions. DETAILED CONTENT STRUCTUREPART 1: DEEP LEARNING FOUNDATIONS (Pages 1-80)Objective: Build mathematical thinking and basic PyTorch skills.Chapter 1: Neural Networks & The Computational EngineLab 1: Build an MLP network to classify handwritten digits (MNIST) using PyTorch.Chapter 2: Optimization & BackpropagationLab 2: Code the Backpropagation algorithm manually (from scratch, without using loss.backward()).Chapter 3: Computer Vision (CNN)Architecture: Pooling Layers, VGGNet (Depth), ResNet (Skip Connections: $y = F(x) + x$).Techniques: Batch Normalization, Data Augmentation.Assignment 1: Transfer Learning with ResNet50 for classifying lung X-Rays.Chapter 4: Classical NLPEmbeddings: Why is One-hot bad? Introduction to Word2Vec (CBOW/Skip-gram). Vector example: "King" - "Man" + "Woman" = "Queen".RNN Family: Recurrent structure, long-term memory issues. Gate structures in LSTM & GRU.Seq2Seq: Basic Encoder-Decoder mechanism.PART 2: THE ERA OF TRANSFORMERS & LLMs (Pages 81-160)Objective: Master the core technologies of ChatGPT and Generative AI.Chapter 5: Attention Mechanism & TransformerLab 3: Code a Self-Attention layer from scratch.Chapter 6: Encoder Models (BERT)Lab 4: Fine-tune PhoBERT to classify Shopee comments.Chapter 7: Decoder Models (GPT) & GenAILab 5: Build a mini-GPT to generate "Luc Bat" poetry (Vietnamese 6-8 verse).Chapter 8: Fine-Tuning & Optimizing LLMs (SOTA)Assignment 2: Admissions consultancy chatbot for R.I.S.E. using RAG + Llama 2.PART 3: DEPLOYMENT & MLOPS (Pages 161-200)Objective: Take the model from Notebook to Production.Chapter 9: Bringing the Model to the WorldLab 6: Dockerize an English-Vietnamese translation application.Chapter 10: Optimization & ScalingQuantization: Compress model from FP32 to INT8.Cloud: Deploy to Google Cloud Run/AWS SageMaker.Distributed: Concept of multi-GPU training.APPENDIX: CAPSTONE PROJECTTopic: All-powerful AI Virtual Assistant.Process: Data Collection -> Fine-tune (Mistral/Llama) -> RAG -> Web App (Streamlit).