Interactive · In-browser training · No dependencies

Neural Networks,
Made Visible

Two interactive experiments that let you build, configure, and train neural networks in real time — and actually watch what they're learning.

gradient-descent backpropagation tanh activation MSE loss universal approximation regression 3D surface fitting
input x
1 input → hidden layers → 1 output
Experiment 01

2D Regression

Configure hidden layers and neuron counts, drag sliders to tune every weight and bias manually, or hit Train to watch gradient descent fit a noisy sine wave in real time. Residual lines visualize MSE live.

tanh hidden linear output MSE loss full-batch GD live training
Open Experiment →
x₁ × x₂ → y surface
2 inputs → hidden layers → 1 output
Experiment 02

3D Surface Fitting

The network now takes two inputs and predicts a surface over the (x₁, x₂) plane. Rotate the interactive 3D plot to see residuals from every angle. Train to watch a flat surface morph into a rippled sin·cos target.

tanh hidden linear output MSE loss full-batch GD rotatable 3D
Open Experiment →
Architecture reference
Hidden activation
tanh(z)
Output activation
linear (none)
Loss function
MSE
Optimizer
full-batch GD
Weight init
Xavier-ish
Max hidden layers
4
Max neurons/layer
10
Training target
sin(x)·cos(x₂)