Pre-Stanford CS224N: NLP with Deep Learning Prep

Full curriculum

7 modules · ~12 hours total

1. Module 1

   Math you actually need (not the math textbook version)

   Refresh just the calculus, linear algebra, and probability that CS224N uses.

   120 min
   • Vectors, matrices, and the dot product as similarity
   • Matrix multiplication shapes and broadcasting rules
   • Partial derivatives and the chain rule, drawn step by step
   • Probability basics: joint, conditional, expectation, KL divergence
   • Why softmax + cross-entropy is the 'natural' loss for classification

2. Module 2

   Python + NumPy for ML, fast

   Get fluent with the exact NumPy patterns CS224N assignments use.

   90 min
   • Vectorised operations and why loops are forbidden
   • Broadcasting, axis arguments, and reshape pitfalls
   • Implementing softmax and cross-entropy from scratch
   • Manual gradient checks (the trick CS224N uses to grade you)

3. Module 3

   PyTorch in 90 minutes

   From `pip install torch` to a trained classifier without magic.

   100 min
   • Tensors, devices, and dtype gotchas
   • Autograd: how `.backward()` actually works
   • nn.Module, optimisers, and the canonical training loop
   • DataLoader, batching, and shuffling correctly

4. Module 4

   Neural network fundamentals (the CS224N starting line)

   Build the mental model lecture 1 of CS224N assumes you already have.

   100 min
   • Logistic regression as a one-layer network
   • Multi-layer perceptrons and the role of non-linearities
   • Initialisation, learning rate, and what 'training instability' looks like
   • Overfitting, regularisation, dropout — diagnosed visually

5. Module 5

   Words as vectors — word2vec and GloVe

   Understand the embedding intuition CS224N spends two lectures on.

   70 min
   • The distributional hypothesis in one sentence
   • Skip-gram and negative sampling, with code
   • GloVe in 10 minutes
   • Why cosine similarity, not Euclidean

6. Module 6

   Sequence models you'll meet in week 2

   Get RNN/LSTM intuition before lecture 5 hits.

   80 min
   • Why RNNs exist (and where they fail)
   • Vanishing gradients, drawn
   • LSTM/GRU as 'gated memory'
   • Bridging to attention — the one paragraph that explains transformers

7. Module 7

   Capstone: a sentiment classifier you actually understand

   Tie it together with a small PyTorch project, graded by a rubric.

   160 min
   • Load and tokenise a real dataset
   • Train, evaluate, and explain the confusion matrix
   • Compare embeddings: random vs pre-trained
   • Write a 1-page report — the same format CS224N expects