Edlan Lab
Experiments in how AI actually works
Two interactive pieces that run entirely in your browser. Watch a language model choose its words, then draw something and watch a spiking neural network recognise it.
Experiment 01
Watch it think
A language model writes one small piece of a word at a time. At every step it weighs several options and picks one. Here you can see the options it nearly chose, how sure it was, and what happens when you send it down a different path.
A recorded example. Load the model to grow your own and explore other paths.
Experiment 02
Draw, and watch it see
Draw a shape and a real neural network recognises it, live, as you draw. Your strokes become a grid of light, spikes race through two layers of artificial neurons, and the possible answers compete for votes until one wins. Then teach it a shape of your own from just three examples.
Watching a demo. Draw your own on the pad: a circle, triangle, square, star, heart, lightning bolt, spiral, arrow, tick or infinity sign.
How it works
The network was trained on thousands of hand-drawn-style doodles of ten shapes, generated by computer with random wobble, tilt and stroke width so that it copes with real, imperfect drawing. It has 784 inputs (a 28 by 28 grid), two hidden layers of 64 and 32 neurons, and ten outputs, and it scores over 99% on doodles it has never seen.
It was then converted into a spiking neural network, the kind brain-inspired chips run. Each lit cell of your drawing fires spikes at a rate set by its brightness; each neuron gathers incoming spikes and fires when it crosses a threshold. The answer is simply whichever output neuron fires most. As the spiking version, it keeps the same accuracy while only doing work when something actually happens.
Teaching a new shape uses few-shot learning: the network records the pattern of activity your three examples create in its last hidden layer, and a new output neuron fires when it sees that pattern again. Everything runs on your device; your drawings are never sent anywhere.
Running visual recognition models on brain-inspired processors was the subject of the PhD research Edlan was founded on.