Neural network Demo
A small network learns to separate two classes using only the raw coordinates. Change its depth, width and activation to see what it can and cannot represent.
Epoch0
Parameters0
Training loss0
Test accuracy0
Filled dots train the network; rings are held out to test it. When the blue test loss climbs while training loss falls, the network is memorising noise.
Data
Network
What it models
A fully connected network: two inputs (the point's coordinates), one to four hidden layers, and a sigmoid output giving the probability of the orange class. 300 points train it; 100 held-out points, drawn as rings, measure how well it generalises.
Method
- Glorot-scaled random initial weights, full-batch backpropagation written for this page.
- Adam optimiser on binary cross-entropy; three epochs per animation frame.
- The decision field evaluates the network on an 80 × 80 grid.