Introduction to deep learning quiz

Question 1

What is deep learning, and how does it relate to machine learning?

Accepted Answer

A subfield of machine learning using multi-layer neural networks that learn features from data

Answer

A type of database optimized for large files

Answer

A rule-based expert system with no learning

Answer

A method that only works on small datasets without computers

Question 2

What is an artificial neuron (perceptron), and what does it compute?

Accepted Answer

It computes activation(weighted sum of inputs + bias)

Answer

It stores data permanently in memory

Answer

It only multiplies two numbers together

Answer

It randomly outputs 0 or 1 ignoring inputs

Question 3

Why are activation functions necessary in neural networks?

Accepted Answer

They introduce non-linearity so the network can model complex functions

Answer

They reduce the number of parameters automatically

Answer

They store the training data inside the model

Answer

They are only used to speed up matrix multiplication

Question 4

What is the ReLU activation function, and why is it popular?

Accepted Answer

f(x) = max(0, x)

Answer

f(x) = 1/(1+e^-x)

Answer

f(x) = x^2

Answer

f(x) = -x for all x

Question 5

What is the sigmoid activation function and one of its drawbacks?

Accepted Answer

It outputs values in (0,1) but suffers from vanishing gradients

Answer

It outputs any real number with no saturation

Answer

It is identical to ReLU for negative inputs

Answer

It can only output exactly 0 or 1

Question 6

What is a feedforward neural network?

Accepted Answer

A network where data flows one way from input to output with no loops

Answer

A network where outputs feed back into inputs in cycles

Answer

A network that only has one neuron

Answer

A network that cannot have hidden layers

Question 7

What roles do weights and biases play in a neural network?

Accepted Answer

Weights scale input importance; bias shifts the activation

Answer

Weights store the dataset; bias counts the layers

Answer

Both are fixed constants that never change

Answer

Bias scales inputs while weights are always 1

Question 8

What is a loss (cost) function in deep learning?

Accepted Answer

A function measuring the difference between predictions and true values

Answer

A function that counts the number of neurons

Answer

A function that randomly initializes weights

Answer

A function that only runs at inference time

Question 9

What is gradient descent and how does it train a neural network?

Accepted Answer

An algorithm that updates weights opposite to the loss gradient

Answer

An algorithm that increases the loss each step

Answer

A way to add more layers automatically

Answer

A method to delete unused neurons

Question 10

What is the role of the learning rate in training?

Accepted Answer

It controls the step size of weight updates

Answer

It sets the number of hidden layers

Answer

It determines the batch size automatically

Answer

It decides which activation function is used

Question 11

What is backpropagation?

Accepted Answer

An algorithm using the chain rule to compute gradients backward through the network

Answer

A way to feed data forward only

Answer

A method to randomly shuffle the dataset

Answer

A technique to compress the trained model

Question 12

What is an epoch in the context of training a neural network?

Accepted Answer

One full pass through the entire training dataset

Answer

A single neuron's computation

Answer

The number of layers in a network

Answer

A type of activation function

Question 13

What is a batch (mini-batch) and why is mini-batch training used?

Accepted Answer

A subset of data processed before each weight update, balancing speed and stability

Answer

The entire dataset always processed at once

Answer

A single layer of the network

Answer

The final output of the model

Question 14

What is overfitting in deep learning?

Accepted Answer

The model memorizes training data and generalizes poorly to new data

Answer

The model is too simple to learn the data

Answer

The model trains faster than expected

Answer

The model has too few parameters

Question 15

What is underfitting and what typically causes it?

Accepted Answer

The model is too simple and performs poorly on both training and test data

Answer

The model memorizes the training data perfectly

Answer

The model has too many layers

Answer

The model only fails on the training set

Question 16

What is dropout and how does it help neural networks?

Accepted Answer

Randomly deactivating neurons during training to reduce overfitting

Answer

Permanently removing layers from the network

Answer

Increasing the learning rate over time

Answer

Adding more data to the training set

Question 17

What is the difference between training, validation, and test datasets?

Accepted Answer

Training fits parameters, validation tunes hyperparameters, test gives final unbiased evaluation

Answer

All three are used to fit parameters equally

Answer

The test set is used to update weights

Answer

Validation data is never used at all

Question 18

What is a convolutional neural network (CNN) primarily used for?

Accepted Answer

Processing image and grid-like data using convolutional filters

Answer

Storing text in databases

Answer

Replacing the loss function

Answer

Generating random numbers

Question 19

What does a convolutional filter (kernel) do in a CNN?

Accepted Answer

A small weight matrix that slides over the image to detect local patterns

Answer

A function that picks the learning rate

Answer

A layer that flattens the output

Answer

A method to shuffle pixels randomly

Question 20

What is the purpose of a pooling layer in a CNN?

Accepted Answer

It downsamples feature maps to reduce size and add invariance

Answer

It adds more neurons to each layer

Answer

It increases the image resolution

Answer

It computes the final loss

Question 21

What is a recurrent neural network (RNN) and what data is it suited for?

Accepted Answer

A network with loops, suited for sequential data like text and time series

Answer

A network only for static images

Answer

A network without any hidden state

Answer

A network that cannot remember past inputs

Question 22

What problem do LSTM networks address compared to vanilla RNNs?

Accepted Answer

They handle long-term dependencies using gates and a cell state

Answer

They make networks shallower

Answer

They remove the need for activation functions

Answer

They only work on image data

Question 23

What is the vanishing gradient problem?

Accepted Answer

Gradients shrink toward zero in deep layers, slowing or stopping learning

Answer

Gradients grow infinitely large and crash training

Answer

Weights become exactly 1 in all layers

Answer

The loss function disappears

Question 24

What is the softmax function and where is it used?

Accepted Answer

It turns scores into a probability distribution summing to 1 for multi-class output

Answer

It is used only in regression problems

Answer

It removes negative weights

Answer

It computes the gradient directly

Question 25

Why is cross-entropy commonly used as the loss for classification?

Accepted Answer

It penalizes confident wrong predictions and pairs well with softmax

Answer

It is only valid for regression

Answer

It ignores the predicted probabilities

Answer

It always equals the mean squared error

Question 26

What is the difference between parameters and hyperparameters?

Accepted Answer

Parameters are learned from data; hyperparameters are set before training

Answer

Both are learned automatically by backpropagation

Answer

Hyperparameters are the weights of the network

Answer

Parameters are chosen by the user manually

Question 27

What is the universal approximation theorem, in simple terms?

Accepted Answer

A sufficiently large network can approximate any continuous function

Answer

Neural networks can only fit straight lines

Answer

Deep networks cannot approximate functions

Answer

Only infinite layers can learn anything

Question 28

Why is non-random weight initialization important in deep networks?

Accepted Answer

Proper initialization avoids vanishing/exploding gradients and breaks symmetry

Answer

Initialization has no effect on training

Answer

All weights should always start at zero

Answer

Initialization sets the learning rate

Question 29

What is batch normalization and what benefit does it provide?

Accepted Answer

It normalizes layer inputs per mini-batch to stabilize and speed up training

Answer

It deletes neurons randomly

Answer

It reduces the number of epochs to one

Answer

It replaces the loss function

Question 30

What does it mean that deep learning learns 'hierarchical features'?

Accepted Answer

Early layers learn simple features and later layers combine them into complex ones

Answer

All layers learn the exact same features

Answer

Only the last layer learns anything

Answer

Features must be hand-designed before training

Test yourself