Skip to content

2. Watching a Network Learn

In Unit 2, clicking Train Model happened in a few seconds, mostly hidden from view. In Playground, you can watch the exact same process happen one step at a time.

What to actually watch

  • Click the play button. An epoch counter starts climbing — one epoch is one full pass through the training data.
  • The big colored panel (the decision boundary) reshapes gradually as training continues — it starts messy or arbitrary, and settles into a boundary that separates the two colors of dots.
  • The loss graph, top right, trends downward as training continues — loss is just a number measuring how wrong the network's predictions currently are. Lower is better.

This is Unit 2, Lesson 2's training step, slowed down

When you clicked "Train Model" in Teachable Machine, this exact process happened — the decision boundary (in that case, a much more complex one, judging images) reshaped itself gradually, and a loss number dropped, exactly like this. Teachable Machine just does it fast enough that you never see the individual steps.

Try it

On the Circle dataset, press play and watch until the loss graph flattens out. Pause, then describe in your own words how the decision boundary changed from the very start to where it ended up.

Tip

If the decision boundary and the loss curve both stop changing, training has converged — continuing to train further usually won't help much from that point, and risks the overfitting problem in Lesson 4.

Next: Learning Rate and Epochs →