Watch It Learn
Watch a net learn, step by step.
What you'll learn in this level
- What an epoch and what the loss mean during training
- How a network gets a little better with every round
- Why noisy data can never be matched perfectly
How a Net Practises
Here you watch a network learn live for the first time. You pick a dataset — points in a spiral or in two rings, for example — and start the training. Then you watch the network separate the points better with every round.
Two numbers run alongside. The epoch counts how many times the network has worked through all the practice examples. The loss says how far off the network currently is — the smaller, the better. When the loss falls, your network is learning.
- 1
Look at the examples
The net gets the practice examples together with the correct answers.
- 2
Guess and compare
It predicts something for each example. Then it compares that with the correct answer.
- 3
Calculate the loss
All the mistakes are turned into a single number: the loss.
- 4
Nudge the weights
Every weight is shifted a little in the right direction. That was one epoch.
Guess first, then experiment
Your training data is a bit shaky, so it is noisy. Can the loss drop to exactly 0 while training?