Module 1 · Foundations revision

Revision guide · ~30 min · dense bullets, not full lessons

Dense bullets for chapters 1.1–1.3. Open full lessons when a bullet feels fuzzy.

1.1 AI & Deep Learning Essentials

Stack of ideas

• AI — goal-directed behavior under uncertainty (rational agent: observe → act → succeed).
• ML — learn parameters from data instead of hand-coding every rule.
• Deep learning — ML with multi-layer neural nets that learn hierarchical features.
• GenAI — models that sample new content (text, images, code) from a learned distribution.

Classical vs learned

• Rules / search / expert systems scale poorly in noisy high-dimensional worlds (vision, speech, open language).
• Supervised learning: labeled (x, y). Unsupervised: structure in x. RL: actions + reward.

Representation

• One-hot — sparse, no similarity. Embeddings — dense vectors; nearby = similar meaning/use.
• Word2Vec / GloVe intuition: “you shall know a word by the company it keeps.”

Neuron → MLP

• Linear: z = w·x + b. Nonlinearity (ReLU, sigmoid, tanh) makes depth useful.
• MLP = stacked linear + activation. Universal approximator in theory; practice needs data, regularization, right capacity.
• Backprop = chain rule through the graph. Gradient descent: θ ← θ − η ∇L.
• Overfit vs underfit; bias–variance tradeoff. Fix overfit: more data, dropout, weight decay, early stop. Fix underfit: capacity / train longer / better features.

Quick self-check

• Can you place ChatGPT inside AI > ML > DL > GenAI?
• Why does ReLU help vanishing gradients vs deep sigmoids?

1.2 Deep Learning Essentials

Training loop

• Batch → forward → loss → backward → optimizer step. Epoch = one pass over data.
• Learning rate too high → diverge; too low → crawl. Schedules and warmups matter for big models.
• Train / val / test split. Never tune on test.

CNN (vision intuition)

• Convolution shares filters spatially; translation equivariance; local receptive fields.
• Pooling / strides reduce spatial size. Depth stacks simple edges → parts → objects.

RNN / LSTM (sequence intuition)

• Hidden state carries past tokens. Vanilla RNNs struggle with long dependencies (vanishing/exploding gradients).
• LSTM/GRU gates control what to keep/forget — still sequential → hard to parallelize → transformers win for most NLP.

1.3 Transformer Architecture

Why transformers

• Parallel over sequence (vs RNN step-by-step). Attention = soft lookup of relevant context.
• Scales with data + compute; foundation of modern LLMs.

Self-attention (Q, K, V)

• For each position: query asks “what do I need?”, keys answer “what do I contain?”, values carry content.
• Scores: softmax(QKᵀ / √d_k) V. Scale by √d_k so softmax stays sharp as d grows.
• Multi-head: several subspaces in parallel (syntax vs semantics, etc.).

Positional encoding

• Attention is permutation-invariant without positions. Add / inject position (sinusoidal, learned, RoPE in modern LLMs).

Decoder stack (GPT-style)

• Causal mask — position i cannot attend to > i (autoregressive next-token training).
• Cross-attention (encoder–decoder MT models): decoder queries encoder keys/values. Decoder-only LLMs skip a separate encoder.

Autoregressive decoding

• Train: predict next token given left context. Infer: sample / greedy / beam / nucleus (top-p) / temperature.
• Temperature ↑ → flatter distribution (more random). top-p keeps smallest set of tokens with cumulative prob ≥ p.

Must-remember diagram in your head

tokens → embed+pos → [N × (MHA + FFN + residual + LN)] → logits → next token

30-minute drill

1. Explain attention to a rubber duck without jargon for 2 minutes.
2. Sketch one transformer block (residual + LN placement).
3. Name three failure modes of deep nets (overfit, vanishing grad, poor LR) and one fix each.