The Modules
20 / 20 ONLINEEach module is a self-contained interactive notebook. Drag, scrub, click anything. Every visualization runs locally in your browser — real math, no canned animations.
A live laboratory for convolutions, kernels, embeddings & neural architectures. Every concept is a slider, every tensor a shape you can rotate, every gradient an animation. No tutorials. No videos. Just experiments.
Each module is a self-contained interactive notebook. Drag, scrub, click anything. Every visualization runs locally in your browser — real math, no canned animations.
Rotate, multiply, broadcast. Watch shapes cascade through dimension flips and dot products in real time.
Scrub a 3×3 kernel across procedural imagery. Sobel, Laplacian, custom — see every multiply & accumulate.
From edges to eyes. Walk the hierarchy of representations from layer 1 through layer 18.
The non-linear bend that makes deep nets work. Plot f(x) and its gradient; compare ReLU, GELU, sigmoid & more.
The number training shrinks. MSE vs MAE vs Huber; cross-entropy vs focal. Drag the error, watch the penalty.
SGD, Momentum, Adam & friends race down bowls, ravines and Rosenbrock. See who reaches the bottom first.
Stop a model memorizing. Live overfitting demo, L1 vs L2 geometry, dropout, and the full toolbox.
One full training step with every number shown: forward, loss, backprop chain rule, weight update. The atom of deep learning.
Drag layers in, watch tensor shapes propagate. Compare your design against ResNet, VGG, MobileNet.
From RNNs to Q/K/V to ViT. The full block anatomy, then hover patches to watch them attend.
Two encoders, one space. Watch the contrastive matrix train, then classify zero-shot over six classes.
No labels. Student matches teacher across crops — and attention maps segment objects on their own.
A forger and a detective in a duel. A real 2D adversarial game — train it live, watch mode collapse.
Threshold, contour, morph. The classical CV toolbox as one big set of sliders.
Flips, jitters, cutout, mixup. Build an augmentation pipeline and preview a batch.
Watch loss curves descend, gradients propagate, decision boundaries form. Tune LR, optimizer, epochs.
Watch a network learn neuron by neuron — live weights, activations, gradient pulses and a forming boundary.
The engine under CLIP & DINO, with normal backbones. Contrastive, triplet, emergent clusters.
Fly through 3D latent space. Cluster, interpolate, find nearest neighbors. PCA / t-SNE / UMAP.
Chain everything. A visual node graph where the output of any module becomes the input of any other. Build pipelines that ingest an image, push it through edges → features → attention → embedding, and visualize every intermediate tensor.
| FOUNDATIONS | 8 modules |
| ↳ matrix.ops | 12 demos |
| ↳ convolution.live | 8 demos |
| ↳ feature.hierarchy | 5 demos |
| ↳ activation.plot | 9 demos |
| ↳ loss.plot | 8 demos |
| ↳ optimizer.race | 6 demos |
| ↳ regularization | 7 demos |
| ↳ mlp.trace | 8 steps |
| ARCHITECTURES | 5 modules |
| ↳ cnn.builder | 9 demos |
| ↳ attention.map | 6 demos |
| ↳ clip.contrastive | 5 demos |
| ↳ dino.selfsup | 5 demos |
| ↳ gan.adversarial | 6 demos |
| LABORATORY | 7 modules |
| ↳ opencv.toolbox | 11 demos |
| ↳ augmentation.pipe | 7 demos |
| ↳ training.sim | 6 demos |
| ↳ training.mlp | live |
| ↳ representation.lab | 4 demos |
| ↳ latent.space | 5 demos |
| ↳ sandbox.graph | ∞ demos |