SML · neural networks

Neural network visualizations

Click-step visual explainers built for SML / FS 25–26.

Four self-contained, browser-only visualizations. Each one is a small interactive deck — click next (or use the arrow keys) to advance one step at a time. Light/dark theme toggles in the top-right of every page.

Building intuition

Stack neurons, build curves

Take one neuron, then two, then many: see how a stack of simple ReLU units composes into the curves a small network can actually represent.

demo2-stack-neurons/
Optimization

Gradient descent

Watch a parameter walk down a loss surface. Step size, momentum, and the high-D loss landscape — without the equations getting in the way.

gradient-descent-viz/
10 scenes · CNNs

Convolutional networks: a deep dive

From "a filter is a little picture" through receptive fields, handcrafted vs. learned filters, and what individual neurons want to see — ending with segmentation as the same machinery, per pixel.

cnn-deepdive/
14 scenes · U-Net · segmentation

U-Net: a deep dive

Big picture → details → big picture. Architecture overview, the three-stage flow through the bottleneck, inside the encoder (conv-blocks + pooling) and decoder (with skip connections), the transposed-convolution mechanic up close, then loss, training, and where the model breaks.

unet-deepdive/