📚 Learning Plan

Week 0: Overview of ML Fundamentals

The ML fundamentals section introduces model evaluation, classical algorithms and more, which becomes the building blocks of the topics in the weeks following.

Week 1-2: Probability Foundations + Markov Assumption

• Probability_and_Markov_Overview
• Topics:
  • conditional_probability_and_bayes_rule
  • naive_bayes_and_gaussian_naive_bayes
  • joint_and_marginal_distributions
  • Markov Assumption: what it is and why it matters in NLP
• Resources:
  • StatQuest: Conditional Probability (YouTube)
  • StatQuest: Bayes' Rule
  • 3Blue1Brown: Bayes theorem, the geometry of changing beliefs
  • StatQuest: Naive Bayes
  • StatQuest: Gaussian Naive Bayes
  • Khan Academy - Probability & Statistics
  • Speech and Language Processing by Jurafsky & Martin Ch. 3 (Markov models)

Week 3: N-gram Models & Language Modeling

• Ngram_Language_Modeling
• Topics:
  • What is an n-gram?
  • How n-gram language models work
  • Perplexity and limitations of n-gram models
• Activities:
  • Implement a bigram/trigram model on a toy corpus
• Resources:
  • The Illustrated Transformer - start with n-gram part
  • Happy-LLM intro chapter
  • Optional: n-gram language model notebook

Week 4: Intro to Information Theory

• Information_Theory

• Topics:

  • Entropy, Cross-Entropy, KL Divergence
  • Why they matter in language modeling
• Activities:
  • Manually compute entropy of a simple probability distribution
  • Implement cross-entropy loss
• Resources:
  • 3Blue1Brown – But what is entropy?
  • Stanford CS224n Lecture 1

Week 5-6: Linear Algebra for ML

• Linear_Algebra_for_ML
• Topics:
  • Vectors, Matrices, Matrix Multiplication
  • Dot product, norms, projections
  • Eigenvalues & Singular Value Decomposition (SVD)
• Activities:
  • Practice via small matrix coding problems (NumPy or PyTorch)
• Resources:
  • 3Blue1Brown: Essence of Linear Algebra
  • Stanford CS229 Linear Algebra Review

Week 7: Calculus + Gradient Descent

• Calculus_and_Gradient_Descent
• Topics:
  • Partial Derivatives
  • Chain Rule
  • Gradients and optimization intuition
• Activities:
  • Derive gradients of simple functions
  • Visualize gradient descent in 2D
• Resources:
  • Khan Academy Calculus (focus on multivariable sections)
  • Gradient Descent Visualization (YouTube)

Week 8-9: Neural Networks & Backpropagation

• Neural_Networks_and_Deep_Learning_Overview
• Topics:
  • Introduction to Perceptron Algorithm
  • Structure of a feedforward neural network
  • Activation functions (ReLU, softmax)
  • Backpropagation algorithm
• Activities:
  • Implement a simple NN from scratch (e.g., on MNIST or XOR)
  • Derive gradient of softmax + cross-entropy
• Resources:
  • Michael Nielsen’s NN book: http://neuralnetworksanddeeplearning.com/
  • CS231n lecture on backprop

Week 10: Integration and Project

• Integration_and_Project
• Goal:
  • Build a mini-project combining n-gram + neural net ideas
  • Example: Predict the next word using both n-gram and a small MLP
• Outcome:
  • Review all learned concepts
  • Prepare to transition to Happy-LLM’s transformer section

---

Phase 2: Modern Deep Learning - Attention & Transformers

Week 11-12: Attention Mechanisms

• Attention & Transformers Overview
• Topics:
  • Why attention? The sequence-to-sequence motivation
  • Attention fundamentals: Query, Key, Value framework
  • Attention mathematics: Scaled dot-product attention
  • Attention vs. RNNs: parallelization and long-range dependencies
• Activities:
  • Implement basic attention mechanism from scratch
  • Visualize attention weights on simple sequences
  • Compare attention to fixed-context approaches
• Resources:
  • The Illustrated Transformer - Jay Alammar
  • 3Blue1Brown: Attention in transformers
  • Stanford CS224N: Attention

Week 13-14: Self-Attention & Multi-Head Attention

• Self-Attention Overview
• Multi-Head Attention Overview
• Topics:
  • Self-attention mechanism: attending within a sequence
  • Permutation equivariance and the need for positional encoding
  • Positional encodings: sinusoidal vs. learned
  • Multi-head attention: parallel attention heads
  • What do different attention heads learn?
• Activities:
  • Implement self-attention from scratch (NumPy or PyTorch)
  • Implement multi-head attention
  • Visualize attention patterns across multiple heads
  • Add positional encodings and observe effects
• Resources:
  • Attention? Attention! - Lilian Weng
  • The Annotated Transformer - Harvard NLP
  • Original paper: Attention is All You Need

Week 15-16: Transformer Architecture

• Transformer Architecture Overview
• Topics:
  • Full transformer architecture: encoder and decoder stacks
  • Encoder: multi-head self-attention + feed-forward networks
  • Decoder: masked self-attention + cross-attention + feed-forward
  • Residual connections and layer normalization
  • Training transformers: learning rate warmup, label smoothing
• Activities:
  • Implement a complete transformer from scratch
  • Train on a small machine translation task
  • Experiment with different hyperparameters (heads, layers, dimensions)
  • Analyze attention patterns in trained model
• Resources:
  • Andrej Karpathy: Let's build GPT
  • Transformer Explainer - Interactive visualization
  • BertViz - Attention visualization tool

Week 17-18: BERT and GPT - Modern Applications

• BERT & GPT Overview
• Topics:
  • BERT (encoder-only): Masked language modeling, bidirectional context
  • GPT (decoder-only): Causal language modeling, autoregressive generation
  • Pre-training vs. fine-tuning paradigm
  • Transfer learning with transformers
  • Prompt engineering and few-shot learning (GPT-3)
  • Instruction tuning and RLHF (ChatGPT, InstructGPT)
• Activities:
  • Fine-tune a pre-trained BERT model for text classification
  • Generate text with GPT-2/GPT-3 using different prompting strategies
  • Compare encoder-only vs. decoder-only architectures
  • Experiment with prompt engineering for various tasks
• Resources:
  • BERT paper - Devlin et al.
  • GPT-2 paper - Radford et al.
  • GPT-3 paper - Brown et al.
  • Hugging Face Transformers Course
  • Stanford CS224N: Pre-training

Week 19-20: Advanced Topics & Integration Project

• Topics:
  • Efficient transformers: Sparse attention, Linformer, Reformer
  • Long-context models: Relative position encoding, ALiBi
  • Vision transformers (ViT): applying transformers to images
  • Multimodal transformers: CLIP, DALL-E, GPT-4
  • State-space models and alternatives to attention
• Integration Project:
  • Build an end-to-end NLP application using transformers
  • Examples:
    • Question answering system with BERT
    • Text summarization with BART/T5
    • Chatbot with GPT-2 fine-tuning
    • Code generation assistant
• Resources:
  • Vision Transformers paper
  • CLIP paper
  • State Space Models survey