Deep Learning and Neural Networks Practice Test

Choose the correct answer, A, B, C or D.

1. Q1.What is an epoch in neural network training?

   • A.One update of the weights
   • B.One full pass through the training data
   • C.One layer of the network
   • D.One prediction on new data
   Show answer

   Answer: B. One full pass through the training data

   An epoch means the model has seen every training example once; training usually runs for many epochs.

2. Q2.What does a loss function do?

   • A.Stores the training data
   • B.Chooses the network architecture
   • C.Measures how far the model's predictions are from the correct answers
   • D.Speeds up the GPU
   Show answer

   Answer: C. Measures how far the model's predictions are from the correct answers

   Training adjusts the weights to make the loss as small as possible.

3. Q3.What is the batch size?

   • A.The number of layers
   • B.The size of the test set
   • C.The number of epochs
   • D.The number of training examples used for one weight update
   Show answer

   Answer: D. The number of training examples used for one weight update

   Gradients are averaged over a batch before each update. Larger batches need more memory.

4. Q4.Which type of layer is best known for detecting local patterns such as edges in images?

   • A.Convolutional layer
   • B.Recurrent layer
   • C.Embedding layer
   • D.Softmax layer
   Show answer

   Answer: A. Convolutional layer

   Convolutional filters slide across the image and respond to local patterns.

5. Q5.Why is ReLU often preferred to sigmoid in hidden layers?

   • A.It always outputs values between 0 and 1
   • B.It reduces vanishing gradients and is cheap to compute
   • C.It removes the need for training data
   • D.It makes the network linear
   Show answer

   Answer: B. It reduces vanishing gradients and is cheap to compute

   Sigmoid saturates for large inputs, so its gradients become tiny; ReLU does not saturate for positive values.

6. Q6.What happens to dropout when a trained model is used for predictions?

   • A.It drops more units
   • B.It doubles the learning rate
   • C.It is turned off
   • D.It deletes the weights
   Show answer

   Answer: C. It is turned off

   Dropout is a training-time regulariser; at inference every unit is used.

7. Q7.How do residual (skip) connections help very deep networks?

   • A.They remove the need for activations
   • B.They reduce the size of the dataset
   • C.They replace the optimiser
   • D.They let gradients flow through shortcut paths, making deep networks easier to train
   Show answer

   Answer: D. They let gradients flow through shortcut paths, making deep networks easier to train

   Adding a layer's input to its output gives gradients a direct path backwards, easing vanishing gradients.

8. Q8.What does a softmax output layer produce?

   • A.Probabilities across classes that add up to 1
   • B.A single yes/no value
   • C.The input image
   • D.Random numbers
   Show answer

   Answer: A. Probabilities across classes that add up to 1

   Softmax turns raw scores into a probability for each class.

9. Q9.Cross-entropy loss is most commonly used for which kind of task?

   • A.Clustering
   • B.Classification
   • C.Regression of house prices
   • D.Data cleaning
   Show answer

   Answer: B. Classification

   Cross-entropy compares predicted class probabilities with the true class.

10. Q10.Training loss keeps falling but validation loss starts rising. What is a sensible response?

    • A.Train for many more epochs
    • B.Stop training early or add regularisation, because the model is overfitting
    • C.Remove the validation set
    • D.Increase the learning rate a lot
    Show answer

    Answer: B. Stop training early or add regularisation, because the model is overfitting

    Diverging training and validation loss is the classic sign of overfitting.

11. Q11.What is special about the Adam optimiser?

    • A.It uses no gradients
    • B.It only works on images
    • C.It adapts the step size for each weight using running averages of past gradients
    • D.It trains without a loss function
    Show answer

    Answer: C. It adapts the step size for each weight using running averages of past gradients

    Adam combines momentum with per-parameter adaptive learning rates, which makes it a strong default.

12. Q12.What usually happens if the learning rate is far too high?

    • A.Training becomes very slow but stable
    • B.The model stops using its GPU
    • C.The dataset becomes larger
    • D.The loss jumps around or diverges
    Show answer

    Answer: D. The loss jumps around or diverges

    Very large steps overshoot the minimum, so the loss bounces or explodes.

13. Q13.What does an embedding layer do?

    • A.Maps words or IDs to dense vectors that are learned during training
    • B.Compresses images into JPEG files
    • C.Splits data into batches
    • D.Measures accuracy
    Show answer

    Answer: A. Maps words or IDs to dense vectors that are learned during training

    Embedding layers are lookup tables of learned vectors, used for tokens, users, products and more.

14. Q14.Why are GPUs well suited to deep learning?

    • A.They have more storage than CPUs
    • B.They perform many matrix operations in parallel
    • C.They do not need electricity
    • D.They can only run one task at a time
    Show answer

    Answer: B. They perform many matrix operations in parallel

    Neural network training is dominated by large matrix multiplications, which GPUs run in parallel.

15. Q15.What is transfer learning?

    • A.Training a model from random weights on every task
    • B.Moving a model to a different computer
    • C.Starting from a pre-trained model and adapting it to a new task
    • D.Copying another team's test set
    Show answer

    Answer: C. Starting from a pre-trained model and adapting it to a new task

    Transfer learning reuses general features learned on a large dataset, so far less task data is needed.