Explorables
Small interactive essays. Drag a slider, hover a cell, and the idea reveals its mechanism instead of being summarised. A slow archive of mine.
- Image kernels, explained visually.2026
A 3×3 matrix is the engine behind sharpen, blur, emboss, and edge detection, and the first move inside every convolutional neural network. Slid across an image, one pixel at a time.
- Attention, in seven tokens.2026
Self-attention as three lines of arithmetic on a six-word sentence. Click any token to make it the query and watch the dot products, softmax, and weighted value sum fall out.
- Gradient descent, by hand.2026
Four canonical 2D loss surfaces and four canonical optimizers. Drop a starting point, push the learning rate up until things get violent, and race SGD against Adam down the same valley.
- An image, as a sum of waves.2026
The 2D Fourier transform on a 32×32 portrait. Click a frequency cell to see the cosine wave it represents; sweep a low-pass radius and watch the picture lose, or recover, its details.
- An image is sixteen patches wide.2020
The Vision Transformer chops an image into a regular grid of patches, treats them like words, and runs a vanilla transformer over the sequence. Slide patch size; click a patch to see what it attends to.
- R-CNN. The slow ancestor.2014
Object detection, part 1 of 4. Two thousand region proposals, each cropped and pushed through AlexNet. Forty-seven seconds an image. The IoU widget is here too.
- Fast R-CNN. Share the features.2015
Object detection, part 2 of 4. Run the backbone once on the whole image; let proposals pull their features out of one feature map through RoI pooling. Forty-seven seconds becomes two.
- Faster R-CNN. Learn the proposals.2015
Object detection, part 3 of 4. Replace Selective Search with a small convolutional head, the RPN, that predicts offsets to a library of anchor boxes. Detection becomes a single network.
- YOLO. Skip the proposals.2016
Object detection, part 4 of 4. An S×S grid; each cell predicts boxes, confidences, and classes in one forward pass. Non-max suppression cleans up the duplicates.