How neural networks train quiz

Question 1

What does it mean for a neural network to "learn" from examples?

Accepted Answer

It adjusts internal weights so predictions get closer to correct answers

Answer

It memorizes every example word for word and never changes

Answer

A human types in the rules by hand for every case

Answer

It deletes wrong examples until only correct ones remain

Question 2

What is a "weight" in a neural network?

Accepted Answer

A number controlling how strongly one neuron affects the next

Answer

The physical mass of the computer chip

Answer

The number of pictures in the dataset

Answer

The speed of the internet connection

Question 3

What is a loss function and why is it useful during training?

Accepted Answer

A formula measuring how wrong a prediction is

Answer

The amount of memory the network uses

Answer

A way to lose unwanted training data

Answer

The time taken to draw the network diagram

Question 4

In simple terms, what is gradient descent?

Accepted Answer

Adjusting weights in the direction that lowers the loss

Answer

Walking uphill to make the error bigger on purpose

Answer

Randomly resetting all weights every step

Answer

Counting how many neurons the network has

Question 5

What is the role of the learning rate in training?

Accepted Answer

The size of each step when updating the weights

Answer

How fast the screen refreshes

Answer

The number of training examples

Answer

The color of the graph lines

Question 6

What does "forward pass" mean in a neural network?

Accepted Answer

Data flowing from input layer to output to make a prediction

Answer

Deleting the first layer of the network

Answer

Sending the answer back to fix the weights

Answer

Saving the model to a file

Question 7

What is backpropagation, in beginner-friendly terms?

Accepted Answer

Sending error backward to find how each weight should change

Answer

Running the network twice as fast

Answer

Backing up the data to the cloud

Answer

Removing neurons that are not needed

Question 8

Why do neural networks need an activation function?

Accepted Answer

To add non-linearity so the network learns complex patterns

Answer

To turn the computer on

Answer

To delete unused data

Answer

To make the network use less electricity

Question 9

What is one training "epoch"?

Accepted Answer

One full pass through the entire training dataset

Answer

One single example shown to the network

Answer

The moment training stops forever

Answer

The first layer of the network

Question 10

What is a "batch" in neural network training?

Accepted Answer

A group of examples processed together before one weight update

Answer

A single neuron in the network

Answer

The final answer of the network

Answer

A type of computer chip

Question 11

What is overfitting, and why is it a problem?

Accepted Answer

When the network memorizes training data and fails on new data

Answer

When the network learns general patterns very well

Answer

When the network runs out of memory

Answer

When the learning rate is exactly correct

Question 12

What is underfitting?

Accepted Answer

The model is too simple and fails even on training data

Answer

The model is too complex and memorizes noise

Answer

The model has perfect accuracy everywhere

Answer

The model uses too much electricity

Question 13

Why do we split data into training and test sets?

Accepted Answer

To check on unseen data whether the model really generalizes

Answer

To make training take longer on purpose

Answer

Because computers need two copies of everything

Answer

To double the size of the dataset

Question 14

What is a bias term in a neuron, separate from a weight?

Accepted Answer

An extra learnable number that shifts a neuron's output

Answer

An unfair opinion the computer has

Answer

The speed of the network

Answer

A type of training error

Question 15

Why are neural network weights usually set to random values before training?

Accepted Answer

To break symmetry so neurons can learn different features

Answer

So neurons stay identical to each other

Answer

Because zero is impossible to store

Answer

To make the network slower

Question 16

What does it mean when we say training has "converged"?

Accepted Answer

The loss has stopped improving and weights have settled

Answer

The network has been deleted

Answer

The dataset has run out of examples

Answer

The screen has frozen

Question 17

What is a label in supervised learning?

Accepted Answer

The correct answer attached to a training example

Answer

The brand name on the computer

Answer

A sticker placed on the keyboard

Answer

The name of the programmer

Question 18

How is the gradient used to update a weight (the basic update rule)?

Accepted Answer

new = old - learning_rate * gradient

Answer

new = old + learning_rate * gradient

Answer

new = old * gradient only

Answer

new = gradient divided by old weight

Question 19

What is the vanishing gradient problem?

Accepted Answer

Gradients shrink to near zero so early layers barely learn

Answer

Gradients become huge and explode

Answer

The network forgets its name

Answer

The dataset disappears

Question 20

What is a validation set used for during training?

Accepted Answer

Data used to tune settings and detect overfitting during training

Answer

Data used only to make training slower

Answer

The same thing as the training set

Answer

A backup copy stored offline

Question 21

What is the difference between a parameter and a hyperparameter?

Accepted Answer

Parameters are learned; hyperparameters are set by humans beforehand

Answer

They mean exactly the same thing

Answer

Hyperparameters are learned; parameters are set by humans

Answer

Both are chosen randomly by the computer

Question 22

Why is a large, varied dataset usually helpful for training?

Accepted Answer

It helps the network learn general patterns and reduces overfitting

Answer

It only slows the computer with no benefit

Answer

It guarantees zero error always

Answer

It removes the need for a loss function

Question 23

What does "epoch" versus "iteration" mean in training?

Accepted Answer

An iteration is one batch update; an epoch is a full pass over all data

Answer

They are identical terms

Answer

An epoch is one batch; an iteration is the whole dataset

Answer

Both mean one single example

Question 24

How can biased training data lead to an unfair AI model?

Accepted Answer

The model learns and repeats the unfair patterns in the data

Answer

Biased data has no effect on the model

Answer

Bias only affects the computer's speed

Answer

More bias always makes the model fairer

Question 25

What is "training accuracy" and why can it be misleading on its own?

Accepted Answer

Accuracy on the training data, which can hide overfitting

Answer

Accuracy on new unseen data only

Answer

The number of layers in the network

Answer

The speed of the training process

Question 26

What is regularization, in simple terms?

Accepted Answer

Techniques that discourage complexity to reduce overfitting

Answer

A way to make the network bigger and more complex

Answer

A method to delete the test set

Answer

A way to speed up the internet

Question 27

What does the technique called "dropout" do during training?

Accepted Answer

Temporarily turns off random neurons to reduce overfitting

Answer

Permanently deletes random neurons

Answer

Drops the worst training examples

Answer

Stops the training early every time

Question 28

Why does training a neural network usually take many steps instead of one?

Accepted Answer

Because small repeated adjustments reach a good solution without overshooting

Answer

Because computers can only do one thing slowly

Answer

Because one step would use too much electricity

Answer

Because the dataset must be read backwards

Question 29

What is the purpose of normalizing or scaling input data before training?

Accepted Answer

It puts inputs on a similar scale for faster, more stable training

Answer

It makes training slower and unstable

Answer

It deletes important features

Answer

It changes the correct labels

Question 30

What is "early stopping" in training?

Accepted Answer

Stopping training when validation performance stops improving

Answer

Starting training before the data is ready

Answer

Stopping the computer to save power

Answer

Removing the first layer early

Test yourself