Train a rule. Set it alive.

Synaptic learns MNIST through a tied, symmetric 3×3 convolution. Its three learned weights and tanh activation are copied unchanged into Automata's direct Neural CA.

loading MNIST

14×14 MNIST 196 input values

3×3 tied convolution corner · edge · center + tanh

10-way readout MNIST classification

The learned kernel

One parameter is shared by every cell of the same kind.

3 values

Synaptictanh

≡

Automataactivation 1

The 10-way readout learns to name digits, but is not part of the CA. The highlighted convolution is the complete local rule being run at right. Its three internal channels are rendered as one grayscale field.

The transferred automaton

Every generation reapplies the learned MNIST convolution; digit seeds keep evolving until you reset them.

waiting

Direct Neural CA · grayscale

Seed with a held-out digit

Caption: expected → prediction from the selected model

Training phase0 / 78

Train accuracy—

Test accuracy—

Train time—

Preparing handwritten examples…