A weighted decision graph with input, hidden, and output layers. Inputs flow from one layer to the next, each connection carries a weight, and each neuron performs a small local computation combining inputs with a bias and an activation function. Proper neural networks have their weights determined by training with error back-propagation and gradient descent, and typically use the same activation function at each neuron. Nicolò Rebughini's Cometeer recommendation engine accidentally replicated this shape: filter + amplifier stages as hidden layers, coefficients as weights (tuned manually instead of learned), amplification methods as per-neuron activation functions, arbitrary adjustments as biases.