Back to the key question: how does a network go from "guesses at random" to "works"? The idea is simple and repeats over and over:

1. Predict. The network takes an example (a photo) and gives an answer — at first almost random.
2. Compare. We compare the guess to the correct answer. How far off is it? That error is measured by the loss.
3. Adjust. The network nudges its weights a little in the direction that reduces the error. That direction is given by the gradients.
4. Repeat. The same over thousands of examples, many times over. One full pass through all the data is an epoch.

predict → compare (loss) → adjust (gradients) → repeat (epochs)
   ▲                                                    │
   └────────────────────────────────────────────────────┘

With each round the guesses get better on average, and the accuracy number climbs. That is "learning" — no magic, just many small corrections.

Do not worry about the words "loss", "gradients", "epochs" — this is just an introduction. We unpack each one in its own lesson, slowly and with code. Today it is enough to feel the overall rhythm: predict → compare → adjust → repeat.

Next lesson we start from the foundation it all stands on: tensors — PyTorch's data structure, which turns an image into numbers and the network into calculations.