Neural Net Demo 1,210 parameters?Hidden layer: 64 inputs × 16 neurons = 1,024 weights + 16 biases = 1,040 Output layer: 16 hidden × 10 outputs = 160 weights + 10 biases = 170 Total: 1,040 + 170 = 1,210

Test accuracy—
Test loss—

Network

·
Click the input grid to draw a digit.
positive weight negative weight line strength = |weight × incoming activation|

Draw a digit, then click any neuron to see and change the weights feeding it. Every weight and bias starts at 0, so the network knows nothing yet. Set weights by hand or press Train to let it learn from printed and handwriting-style digits; every change is scored live against 1,000 held-out samples.

Training

preparing…

Training set: 4,000 digits rendered from 12 fonts with random tilt, slant, stretch and stroke thickness. Plain gradient descent, sigmoid hidden units, softmax output. Train after editing to watch the network repair your changes.