Learning is not a single correction but a repeated cycle. The model predicts on a batch of examples, the errors are combined into a loss, and each adjustable setting is nudged in the direction that reduces that loss. Then the whole cycle runs again with the updated settings. Because the settings have changed, the next round of predictions is different, so the errors and the loss change too. Over many rounds the loss tends to fall, meaning the predictions drift closer to the correct answers. The size of each nudge matters: too large and the model overshoots and the loss bounces around; too small and progress is painfully slow. Watching the loss curve over rounds is how you tell whether the loop is actually learning.
Learn MotionHow Machines Learn from Data
Measuring Mistakes and Adjusting
The Learning Loop in Motion
3 / 3
Watch the cycle repeat. In each round the model makes a prediction, the gap to the correct answer becomes an error, the errors combine into a loss, and the settings shift slightly in the direction that lowers that loss. Because the settings changed, the next prediction is different, so the next loss is different too. Follow the loss curve as the rounds accumulate: it should trend downward, with each step smaller than the last as the model settles. If the curve bounces or climbs, the adjustment step is too large.
0:00 / 0:00