Stop writing rules; start learning from data. We transition from hand-coded algorithms to the Data-Driven Paradigm. You will start by analyzing the "Lazy" Nearest Neighbor approach, then move to the industry-standard Parametric approach (Linear Classifiers). Finally, you will demystify the "learning" process by breaking down Loss Functions (Cross-Entropy) and the engine of all Deep Learning: Gradient Descent. Topics: Data-Driven vs. Hand-coded, Nearest Neighbor (L1 Distance), Linear Classifiers & Matrix Math, Softmax & Cross-Entropy Loss, Optimization via Gradient Descent. Lab 0: Implement k-NN from scratch + Intro to Linear Scores using NumPy.

Rip the neural net open. Compute forward passes. Derive gradients by hand. Verify with gradient checks. After this week, autograd is convenience - not magic. You'll never be confused by a PyTorch error again. Topics: Computational graphs explained, FC/ReLU/Conv forward passes, backpropagation derivations, gradient checking methodology. Lab 1: Manual backprop in NumPy with gradient verification.

SGD, momentum, Adam - know when to use what. Learning rate schedules that actually work. Weight decay, BatchNorm, LayerNorm. This week separates "ran a notebook" from "trained a model properly." Topics: Gradient Descent / SGD fundamentals, momentum & Adam optimizers, weight decay & L2 regularization, learning rate schedules, initialization strategies, BatchNorm vs. LayerNorm intuition. Lab 2: Optimizer shoot-out - compare training curves.

Tensors, autograd, nn.Module - mastered. Clean training loops you'd ship to production. DataLoaders, hooks, checkpoints. Professional patterns. After this, your code looks like it came from a real ML team. Topics: Tensors & autograd deep dive, nn.Module architecture, Dataset & DataLoader patterns, hooks for debugging/inspection, clean train/eval loop structure. Lab 3: Train a small ConvNet end-to-end in PyTorch.

Trace the evolution. LeNet → AlexNet → ResNet. RNN → LSTM → Attention. Know why each architecture was invented and what it solved. History = pattern recognition for building what's next. Topics: LeNet → AlexNet → VGG → ResNet → Inception, RNN → LSTM → GRU progression, the rise of Attention mechanisms, model zoo timelines. Mini-reading: Choose one landmark paper to skim. Quiz on architecture evolution.

The architecture behind GPT, BERT, everything modern. Self-attention from first principles. Multi-head attention. Positional encodings. Scaling laws. Then you build a character-level Transformer yourself. Small model, massive insight. Topics: Self-attention mechanism unpacked, Multi-Head Attention (MHA), positional encodings (sinusoidal, learned), scaling laws & inference tricks. Lab 4: Build a tiny character-level Transformer on a toy corpus.

Generate images from noise. GANs: the adversarial game, failure modes, fixes. VAEs: ELBO, latent spaces, reconstruction. Understand why training is unstable - and how to stabilize it. You'll generate MNIST/CIFAR images. Topics: GANs (minimax game, failure modes, stabilization tricks), VAEs (ELBO derivation, KL divergence intuition), RNN seq2seq vs. attention-based generation. Lab 5: DCGAN or VAE on MNIST/CIFAR (demo + code).

Where it all converges. Diffusion models: noise-to-image pipelines, U-Net, schedulers. LLMs: tokenization (BPE), embeddings, post-training techniques. SFT, RLHF, GRPO - the methods that turn base models into useful assistants. Capstone begins. Topics: Diffusion intuition (forward/reverse process), U-Net architecture & schedulers, Classifier-Free Guidance (CFG), LLM fundamentals (tokenizers, embeddings, logits→tokens), post-training (SFT, RLHF intuition, GRPO intuition), evaluation & safety. Lab 6: Tokenizer exploration + SFT-style finetune stub. Capstone kickoff.