Library
Machine learning visualisations
Search by concept or filter by topic, then open any workspace directly. There is no required order—start wherever your question is.
How can the same local next step converge, oscillate, diverge, or find a different basin?
Change the start and step size on computed 3D loss surfaces, then compare direct descent, ravine oscillation, divergence, and local-minimum traps.
When does a flexible model stop generalising?
Fit polynomials of increasing degree to noisy samples, then compare training and validation error to see where the model stops generalising.
How does k-means decide where clusters belong?
Move the centroids, step through assignment and update phases one at a time, and watch the inertia settle as the split improves.
How can a feature map turn a circular boundary into a flat one?
Map radial distance to a third coordinate, then connect a flat feature-space threshold to its circular input-space boundary.
How does a self-attention head turn query-key similarity into weights?
Select a query token and inspect the similarity scores and softmax weights it computes against every key.
How do temperature and truncation change what a language model writes?
Adjust temperature and truncation over a hand-authored next-token distribution, then sample to see which continuation gets picked.
How does a CNN turn pixels into features?
Move from an input image through convolution, activation, and pooling while inspecting the exact patch calculation behind every feature-map cell.
How can a swarm find an optimum without gradients?
Watch particles balance momentum, their own best discovery, and the swarm’s shared best while searching a landscape full of local minima.
How does evolution search without knowing a gradient?
Evolve a binary population through selection, crossover, and mutation while tracking fitness and genetic diversity across two competing peaks.
How does PCA compress data without labels?
Rotate a one-dimensional projection through a two-dimensional point cloud and watch retained variance trade against reconstruction error.
How does error travel backward through a neural network?
Follow inputs through a small neural network, measure prediction error, and trace the gradients that assign each weight responsibility before an update.
How do model parameters move a fit or decision boundary?
Move slope and intercept, then connect the visible prediction line to the same coordinates on a linear or logistic loss surface.
How does a decision tree carve up feature space?
Reveal a small tree one level at a time and move its first threshold to see rules become rectangular prediction regions.