Take the Deep Learning — Lecture 2, Set 01: 20 multiple-choice questions (30 minutes) with instant scoring and a full answer review. Covers 20 Deep…
The first 5 questions of this test, with the correct answer and explanation.
Q1. What exactly is MNIST?
- A. 70,000 small grey-scale pictures of the digits 0 to 9, each 28 x 28 pixels, each with a label ✓
- B. 1.4 million colour photographs in 1,000 categories
- C. 60,000 handwritten letters of the English alphabet
- D. A library for building neural networks in Python
Answer: A. 70,000 small grey-scale pictures of the digits 0 to 9, each 28 x 28 pixels, each with a label
Explanation: MNIST is 70,000 grey-scale 28 x 28 pictures of the digits 0 to 9, each with a label saying which digit it is.
Q2. How is MNIST split between learning and testing?
- A. 50,000 for learning and 20,000 for testing
- B. 60,000 for learning and 10,000 kept aside to test ✓
- C. 35,000 for learning and 35,000 for testing
- D. 70,000 for learning, with testing done on the same images
Answer: B. 60,000 for learning and 10,000 kept aside to test
Explanation: 60,000 images are used for learning and 10,000 are kept aside to test.
Q3. How many numbers go into the network for one image, and how many come out?
- A. 28 in and 10 out
- B. 512 in and 10 out
- C. 784 in and 512 out
- D. 784 pixel values in and 10 numbers out ✓
Answer: D. 784 pixel values in and 10 numbers out
Explanation: The network takes the 784 pixel values in and gives out 10 numbers: how likely the picture is each digit from 0 to 9.
Q4. According to the road map, what does learning (training) mean?
- A. Show an example, compare the answer with the true label, measure the error with one number, nudge each weight to shrink it, repeat ✓
- B. Store every training image so the network can look the answer up later
- C. Write explicit rules for each handwriting style and test them
- D. Increase the number of layers until the accuracy stops improving
Answer: A. Show an example, compare the answer with the true label, measure the error with one number, nudge each weight to shrink it, repeat
Explanation: Training means: show an example, compare with the true label, measure the error with the loss, work out which direction to nudge each weight, nudge a little, and repeat tens of thousands of times.
Q5. What two layers make up the MNIST network in the Keras code?
- A. Dense(784, 'relu') and Dense(10, 'sigmoid')
- B. Dense(512, 'softmax') and Dense(10, 'relu')
- C. Dense(512, 'relu') and Dense(10, 'softmax') ✓
- D. Conv2D(512) and Dense(10, 'softmax')
Answer: C. Dense(512, 'relu') and Dense(10, 'softmax')
Explanation: The network adds Dense(512, activation='relu') then Dense(10, activation='softmax').
Take the timed test (free, no sign-up) to see your score and the answer and explanation for every question.
Q6. What do the 512 units in the first Dense layer each produce?
- A. One number from all 784 pixels
- B. One probability for each of the 10 digits
- C. A 28 x 28 feature map
- D. A copy of one of the input pixels
Q7. In compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy']), what do the three arguments choose?
- A. The dataset, the batch size and the number of epochs
- B. The number of layers, the number of units and the activation
- C. How to adjust the weights, how to measure the error, and what to report to us
- D. The hardware, the precision and the random seed
Q8. Why are the pixel values reshaped and divided by 255?
- A. To reduce the file size of the dataset
- B. To flatten each image into one row of 784 numbers and rescale them to 0-1, so small values of similar size keep the arithmetic stable
- C. To convert the images from colour to grey-scale
- D. To remove images that are too dark or too bright
Q9. What does to_categorical do to the label 3, and why?
- A. It converts it to the string "three" for readability
- B. It scales it to 0.3 so it matches the pixel range
- C. It leaves it as 3 but changes its data type to float32
- D. It rewrites it as [0,0,0,1,0,0,0,0,0,0], the same shape as the 10 output probabilities
Q10. In fit(train_images, train_labels, epochs=5, batch_size=128), what is an epoch?
- A. One pass over all 60,000 training images
- B. One batch of 128 images
- C. One weight update
- D. One forward pass through a single image
Q11. What test accuracy does the MNIST example reach, and on what data?
- A. 0.9273 on the training images
- B. 0.9692 on a validation split
- C. 0.9785 on 10,000 images the network has never seen
- D. 1.0000 on the images it was trained on
Q12. A scalar such as np.array(12) is which kind of tensor, and what is its ndim?
- A. A 1D tensor with ndim 1
- B. A 0D tensor with ndim 0
- C. A 2D tensor with ndim 2
- D. Not a tensor at all
Q13. np.array([12, 3, 6, 14]) is:
- A. A scalar (0D tensor)
- B. A matrix (2D tensor)
- C. A 3D tensor
- D. A vector (1D tensor)
Q14. What are tensors, in the lecture's definition?
- A. Multidimensional NumPy arrays, containers for numbers
- B. Functions that transform one layer into the next
- C. The adjustable weights of a neural network
- D. A type of activation function
Q15. What are the three tensor attributes described in the lecture?
- A. Size, speed and precision
- B. Number of axes (rank), shape, and data type (dtype)
- C. Width, height and depth
- D. Weights, biases and activations
Q16. For MNIST, what do train_images.ndim, train_images.shape and train_images.dtype return?
- A. 2, (60000, 784) and float32
- B. 3, (60000, 28, 28) and float64
- C. 3, (60000, 28, 28) and uint8
- D. 4, (60000, 28, 28, 1) and uint8
Q17. Which tensor shape is correct for time-series or sequence data?
- A. 2D tensors of shape (samples, features)
- B. 3D tensors of shape (samples, time-steps, features)
- C. 4D tensors of shape (samples, height, width, channels)
- D. 5D tensors of shape (samples, frames, height, width, channels)
Q18. What is the shape of a batch of 128 colour images of size 256 x 256, in channels-last convention?
- A. (128, 3, 256, 256)
- B. (128, 256, 256, 1)
- C. (256, 256, 3, 128)
- D. (128, 256, 256, 3)
Q19. Which framework pairing matches the two image-tensor conventions correctly?
- A. Channels-last is the TensorFlow and Keras default; channels-first is used by Theano and PyTorch
- B. Channels-first is the TensorFlow and Keras default; channels-last is used by Theano and PyTorch
- C. Both TensorFlow and PyTorch use channels-last
- D. The convention is chosen per image, not per framework
Q20. What does train_images[:, 14:, 14:] select?
- A. Images 14 to the end, in full
- B. The first 14 rows and columns of every image
- C. The 14 x 14 pixels in the bottom-right corner of all images
- D. Every 14th image in the dataset