How do we reduce loss?

Explore how changing weights and biases affects the output—and why a change that helps one example may hurt another.

Open the training activityKnowledge check, introduction, and three practice steps.

One example

Start with zero weights and biases in a small network. Change connections and nodes to bring its output closer to the target.

One network, four examples

The AND target is 1 only when both inputs are 1. Cycle through all four input pairs while keeping the same weights and biases. Compare the average loss, not just one example.

Two outputs

Now the network must match two targets for every input pair. Add the two squared errors for each example, then average the four example losses.

Try a few changes, then load the worked solution. Finding every setting by guessing is not the goal. Meeting 4 will introduce a systematic way to choose adjustments.