ConvNetJS 2D Classification

Question 1

ConvNetJS 2D classification demonstrates a neural network…

Accepted Answer

Training entirely in the browser — It runs a live neural network in JavaScript, letting you watch it learn a decision boundary on 2D data.

Question 2

The colored regions in the demo represent the network's…

Accepted Answer

Predicted class for each point in the 2D plane — Each region's color is the model's prediction there; the boundary between colors is what the network learns.

Question 3

To separate non-linearly arranged points (e.g. circles), the network needs…

Accepted Answer

Hidden layers with non-linear activations — A linear classifier can't separate concentric or interleaved classes; hidden layers create the needed non-linear boundary.

Question 4

As training proceeds, the decision boundary…

Accepted Answer

Gradually reshapes to better separate the classes — Gradient descent updates the weights each step, and you watch the boundary morph to fit the data.

Question 5

Adding MORE neurons/layers lets the boundary become…

Accepted Answer

More complex and flexible — Greater capacity can carve more elaborate boundaries; too much capacity may overfit the training points.

Question 6

A boundary that wiggles to perfectly enclose every training point likely indicates…

Accepted Answer

Overfitting — Overly contorted boundaries generalize poorly; simpler boundaries or regularization usually perform better on new data.

Question 7

The network is trained using…

Accepted Answer

Gradient descent / backpropagation on a classification loss — The browser network computes gradients and updates weights to reduce misclassification, exactly like larger frameworks.

Question 8

ConvNetJS runs without any server because…

Accepted Answer

All computation happens client-side in JavaScript in your browser — Being pure JavaScript, the whole training loop executes locally, making it a convenient interactive teaching tool.

Question 9

With too FEW neurons for the data's shape, the network will…

Accepted Answer

Underfit the training data — Insufficient capacity (high bias) leaves the boundary unable to follow the true class structure.

Question 10

The activation function in the hidden units provides…

Accepted Answer

The non-linearity that lets the network bend the boundary — Without non-linear activations, the network could only draw straight-line boundaries regardless of size.

Train a classifier in the browser

Test your knowledge

ConvNetJS 2D classification demonstrates a neural network…

Interview Questions

References & Further Reading

Love this visualization?