WEBVTT

00:00.000 --> 00:13.000
This is TF Playground, a tool built by Google to let you experiment with neural networks

00:13.000 --> 00:14.000
in your browser.

00:14.000 --> 00:16.920
No login, no download, it runs right here.

00:16.920 --> 00:19.000
The screen has four zones.

00:19.000 --> 00:25.120
Top left, your data set, left column, your input features, center, your network, right,

00:25.120 --> 00:28.800
the output, the decision boundary your network draws.

00:28.800 --> 00:34.200
We will use the circle data set, the two rings, it should already be selected, if not, click

00:34.200 --> 00:35.200
it now.

00:35.200 --> 00:38.040
Now press play, watch the right side of the screen.

00:38.040 --> 00:42.920
The network is training, each tick of that counter is one training round, one forward

00:42.920 --> 00:45.400
pass and one backward pass.

00:45.400 --> 00:49.040
See how the boundary settles into a shape that separates the two rings?

00:49.040 --> 00:51.120
The network has learned to tell them apart.

00:51.120 --> 00:53.520
This is what training looks like from the outside.

00:53.520 --> 00:58.320
Now stop training first, always stop before you change the network, good.

00:58.320 --> 01:02.680
Now look at the center panel, this network has one hidden layer, click the plus button

01:02.680 --> 01:04.680
to add a second hidden layer.

01:04.680 --> 01:07.600
Now press play again, watch the output.

01:07.600 --> 01:10.760
The boundary looks different, more complex.

01:10.760 --> 01:14.560
More layers means the network can learn more complex shapes.

01:14.560 --> 01:18.360
Stop again, remove a layer with the minus button, press play.

01:18.360 --> 01:24.400
The boundary goes back to a simpler shape, fewer layers, simpler patterns, more layers,

01:24.400 --> 01:26.080
more complex patterns.

01:26.080 --> 01:31.520
You are in control of the architecture, three things to notice every time you use playground,

01:31.520 --> 01:34.080
one, the decision boundary.

01:34.080 --> 01:36.280
Does it separate the data correctly?

01:36.280 --> 01:39.680
Two, how many training rounds did it need?

01:39.680 --> 01:42.840
Three, how does the architecture change the result?