Follow Numbers Forward

Prediction is repeated transform and activate

A forward pass sends an input through every layer to produce a prediction. Each dense layer computes weighted sums plus biases, applies an activation, and passes the resulting vector onward. No parameters change during the forward pass itself.

input x = [2, 1]
hidden score = 2×0.5 + 1×(-1.0) + 0.5 = 0.5
hidden ReLU = 0.5
output score = 0.5×2.0 - 0.2 = 0.8
output sigmoid ≈ 0.69

Real networks vectorize these operations across many units and examples using matrices. The principle remains the same: deterministic arithmetic under the current parameter values.

Analogy: A document moves through an approval chain. Each desk reads the current form, adds a transformation, and passes a new form forward. The final desk produces the decision.

Note: Inference is essentially forward passes under fixed parameters. Training adds loss calculation, backpropagation, and parameter updates afterward.