Neural Network Playground (Build & Train)

Question 1

In a feedforward neural network, each neuron computes its output by:

Accepted Answer

Taking a weighted sum of its inputs plus a bias, then applying an activation function — A neuron's output is activation(sum(weight_i * input_i) + bias) - a weighted sum plus bias, passed through a non-linear activation.

Question 2

Why is a non-linear activation function (like sigmoid) necessary between layers?

Accepted Answer

Without it, stacking multiple layers would collapse into an equivalent single linear transformation, unable to learn non-linear patterns — Composing linear functions only ever produces another linear function, no matter how many layers you stack; nonlinearity is what lets deep networks represent curved decision boundaries.

Question 3

The XOR problem is significant in neural network history because:

Accepted Answer

It is a simple pattern whose four cases cannot be separated by any single straight line, which is why hidden layers are necessary — XOR's positive and negative examples can't be separated by any single straight line, which historically showed the limits of the single-layer perceptron and motivated multi-layer networks.

Question 4

Why can a network with zero hidden layers not solve XOR?

Accepted Answer

A single-layer model can only form one straight decision boundary, and XOR's four cases cannot be separated by any single straight line — No amount of training changes the fact that a linear model can only represent one straight decision boundary, and XOR is fundamentally not linearly separable.

Question 5

Adding a hidden layer helps solve XOR because:

Accepted Answer

It lets the network combine multiple linear boundaries into a non-linear decision region — Hidden neurons each learn their own linear boundary, and the output layer combines them nonlinearly, together carving out the curved region XOR requires.

Question 6

In a neural network playground, what do the "depth" and "width" settings control?

Accepted Answer

The number of hidden layers (depth) and the number of neurons per hidden layer (width) — Depth is how many hidden layers the network stacks; width is how many neurons sit in each of those layers.

Question 7

The forward pass in a neural network refers to:

Accepted Answer

Propagating input values layer by layer to produce a final prediction — The forward pass feeds an input through each layer in sequence, computing activations until the output layer produces a prediction.

Question 8

Backpropagation is the algorithm used to:

Accepted Answer

Compute how much each weight contributed to the prediction error, so weights can be updated to reduce that error — Backpropagation applies the chain rule to propagate the output error backward through the network, computing a gradient for every weight.

Question 9

The learning rate controls:

Accepted Answer

The size of the step taken when updating weights after each gradient computation — After computing gradients, the learning rate scales how big a step the optimizer takes in the direction that reduces the loss.

Question 10

What typically happens if the learning rate is set far too high?

Accepted Answer

The loss can oscillate wildly or diverge ("explode") instead of steadily decreasing — Each update step scales with the learning rate, so an oversized rate overshoots the minimum back and forth — the loss oscillates or grows until the weights blow up.

Training Control

Architecture

Hyperparameters

Debug a network that will not learn

Test your knowledge

In a feedforward neural network, each neuron computes its output by:

Interview Questions

References & Further Reading

Love this visualization?