forward pass

The forward pass is simply the act of running data through a neural network from start to finish to get an answer. You hand the network an input — a photo, a sentence, a row of numbers — and it flows through each layer in order: every neuron computes its weighted sum, adds its bias, applies its activation, and passes its output to the next layer, until the final layer produces a result. It's like dropping a question into the top of a machine and watching the answer fall out the bottom.

Nothing is learned during a forward pass. The weights and biases are held fixed; the pass just uses them to compute. This is exactly what happens every time you actually use a trained model — when a translation app renders your sentence, or a photo app tags a face, it is performing a forward pass. In that sense the forward pass is inference: the network applying what it already knows.

During training, the forward pass is only the first half of each step. You run an input forward to get the network's prediction, then compare it to the right answer using a loss function (a measure of how wrong it is). The second half — the backward pass, or backpropagation — traces that error back through the same network to figure out how to adjust each weight. So the forward pass shows up in two roles: on its own during everyday use, and as the opening move of every learning step.