Topics

Interactive deep-dives into ML algorithms. Each topic has a visual story, annotated code, quizzes, and an internals explorer.

Word2Vec

Word Embeddings & Skip-gram

Global Vectors for Word Representation

Build and interpret a word-word co-occurrence matrix with distance weighting
Understand why probability ratios P(k|i)/P(k|j) encode meaning better than raw probabilities
Follow the 5-step derivation from ratios to the log-bilinear model
Explain the weighted least squares objective and why each design choice matters
Implement GloVe training with manual gradients and AdaGrad
Understand WHY vector arithmetic (king − man + woman ≈ queen) works mechanically in GloVe's log-bilinear framework
Connect GloVe to PMI, SVD, LSA, and the Levy-Goldberg result; know when GloVe outperforms alternatives and when it doesn't

Recurrent Neural Networks

Understand why sequences need memory and how hidden state provides it
Build a vanilla RNN from five raw parameter tensors (no nn.RNN)
Implement Backpropagation Through Time (BPTT) manually, line by line
Visualize vanishing gradients and understand why eigenvalues matter
Watch category emergence from character-level prediction (Elman 1990)
Generate text character by character with temperature-controlled sampling
Understand the architectural ceiling of vanilla RNNs and why gating mechanisms were needed