The Shape of Learning
An atlas of learning machines. Every image is drawn by the computation it describes.
A small network learns two interleaved spirals in your browser. Each line is its decision boundary after one training step; light gathers where the boundary stays.
NowPreparing the network.
Learning happens in time. These plates are long exposures of it: each records where a learning system went, and how long it stayed.
Brightness is time
Light accumulates wherever a process lingers: a boundary that settles, a path that converges, an agent that keeps returning.
Ember and ice are two sides
Warm light marks the first class, task, or valley; cool light marks the second. Where they overlap, they add toward white.
Every plate names its medium
Exact computation, designed transformation, trained in your browser, or simulated experience. Each plate says which, and where its model stops.
The plates
Six experiments across machine learning, deep learning, reinforcement learning, and continual learning. Each begins with a question you can test with your own hands.
How the shapes connect
The plates are not chapters. They share mathematics, and each connection below is a statement you can check against two plates.
The project
The Shape of Learning studies how learning works through objects you can move, fields you can reshape, and changes you can trace.
- 01
Machine learning
What makes examples separable, and when does fitting them stop predicting new ones?
- 02
Deep learning
How do learned transformations, optimisers, and generative flows move points and probability?
- 03
Reinforcement learning
How do rewards that arrive later reshape what an agent does now?
- 04
Continual learning
What survives when a system learns something new, and what protects it?
Computed, not drawn
No image on this site is a picture of an idea. Each is rendered from a stated model at the moment you look: a network trained in your browser, an exact solution, or a simulation you can rerun.
Brightness has a meaning. Each plate accumulates light like a long photographic exposure, so a bright region is one the process visited often or for long.
Honest about its medium
Toy models can be exact and still mislead. Every plate states whether it shows an exact computation, a designed transformation, a trained model, or simulated experience, and where its model stops matching practice. Research claims cite their primary sources.
Made to be handled
Every plate starts with a question and a few guided moves: predict, act, observe. All controls work from the keyboard, draggable points are focusable, and reduced-motion preferences replace animation with the final state.
Self-contained
Everything runs locally in the browser with bundled fonts and math. There is no tracking, no server computation, and no network request after the page loads. Press / anywhere to open the index.