Experiment 001 / Machine learning
A network, learning.
Train a tiny neural network. Add your own points. See the pattern it learns.
This experiment needs JavaScript. You can also read the illustrated note below.
The model is untrained. Train it to see a boundary emerge between two classes of points.
Teach a small
network a new trick.
Can 105 numbers learn a pattern? Pick a dataset, press train, and watch the boundary take shape.
- Steps
- 0
- Loss ↓
- —
- Training fit
- —
80 points. Ready when you are.
Under the hood
Two inputs, two hidden layers of eight neurons, one output. Tanh activations, binary cross-entropy loss, and Adam optimization, written in plain JavaScript. The score is fit to these training points, not a measure of general intelligence.
Read the computation source ↗Inside the machine
Follow
the signal.
This is the same network you trained above. Move the inputs and watch the activations change. Blue connections are positive; orange connections are negative.
