The six sessions 🧠 · one network, built and told the truth about
Everyone starts at session 1 with three numbers and a squash. By session 6 you have trained a real network live, tamed it, and measured it against the strongest tabular baseline in the estate.
The neuron
Weights, bias, activation - computed by hand until boring. Why stacking linear things buys nothing, and what the famous theorem refuses to promise.
⛓️ Session 2 · mediumBackprop by hand
The chain rule walked through a scalar graph, micrograd-style: every weight gets its exact share of the blame - plus the honest history of who invented what.
🌀 Session 3 · mediumThe training loop
A real network carves boundaries live: XOR falls exactly as predicted, and the spiral separates capacity from optimization - the lesson no theorem could teach.
👁️ Session 4 · hands-onSeeing and embedding
Convolution as a promise about images, the two-GPU week that changed everything, residual connections that made 152 layers trainable - and embeddings as the product.
🎛️ Session 5 · hands-onTaming the net
Overfit diagnosis, dropout quoted from the source, the get-more-data rule, random labels fitted perfectly - and the double-descent curve that bent the textbook.
⚖️ Session 6 · capstoneWhen deep learning
The showdown: four networks vs the tree course's tuned booster on byte-identical data, live - the best net lands 68-70 against 73, and the modality map explains the loss and the empire at once.
Your passport 🎫
Clear all three quiz questions in a session and it gets stamped. Stamps live in this browser only.
Choose your path 🗺️
Three doors into the same discipline: build it, train it, and tell the truth about it.
What is real here 🧾
A deep learning course owes you unusual precision about what is genuinely computed.
| Claim | Status |
|---|---|
| The networks in the widgets | Real. Genuine forward and backward passes (full-batch gradient descent, real chain rule) run in your browser on every press; the boundary heatmap is the live model queried on a pixel grid |
| The showdown numbers: 64 / 65 / 60 / 68-70 vs 73 | Measured. The generator is copied byte for byte from the ensemble course's engine - same seed, same split - and every network refits live. The deep run lands 68-70 because browsers' float math libraries differ slightly and 1,500 chaotic epochs amplify it - stated on the widget |
| The datasets | The teaching artifice: seeded 2D toys (moons, xor, circle, spiral) and the seeded Mango Lane sample, stated on every widget |
| Every research claim | Verified to source - including the misquote watchlist taught as content: backprop was popularized in 1986 not invented; the approximation theorem proves existence only; AlexNet's 15.3 belonged to the ensemble entry; AdamW, not Adam, is the modern default |
| Trees vs nets on tabular | Scoped honestly per Grinsztajn et al.: typical, medium-sized data - and the course reproduces it live rather than asserting it |
| The in-browser engine vs PyTorch | A teaching implementation - the five-move loop is identical; minibatches, Adam and torch.compile are what production adds, covered in session 3 |
The knowledge map 🧠
The whole course at a glance - hover a session to spotlight its concepts, click any node to jump in.