QUIZ
CNNs and vision
9 questions on CNNs and vision, from 9 lessons. Pick an answer in your head, then open it.
On the roadmap for: ML Engineering.
Question 1 of 9
Why does a 3×3 convolution suit images better than a dense layer?
- A. It ignores neighbouring pixels to avoid picking up noise
- B. It's always more accurate on every kind of data, not just images
- C. It only works on grayscale images with one channel
- D. It shares local weights across positions, so far fewer parameters
View answer
✓ D. It shares local weights across positions, so far fewer parameters
Locality and weight sharing match how images work: an edge looks the same anywhere.
Lesson: Convolution and Receptive Field
Question 2 of 9
Input 32×32, kernel 3×3, padding 1, stride 2. What is the output's height and width?
- A. 32×32
- B. 30×30
- C. 16×16
- D. 15×15
View answer
✓ C. 16×16
Output = ⌊(32 − 3 + 2·1) / 2⌋ + 1 = ⌊31/2⌋ + 1 = 16.
Question 3 of 9
Before ResNet, why did much deeper plain networks get worse results, even on training data?
- A. Too little training data for such deep models to learn anything from
- B. GPUs ran out of memory for deeper models
- C. Harder optimisation: signals faded through many layers
- D. Overfitting the training data with too many layers
View answer
✓ C. Harder optimisation: signals faded through many layers
Higher training error showed it was an optimisation problem; skip connections fixed it.
Lesson: ResNets and Skip Connections
Question 4 of 9
What idea did VGG add to CNN design?
- A. Skip connections around each block of layers
- B. Attention layers between the convolution blocks
- C. Stacking small 3×3 convolutions for a big receptive field
- D. Depthwise separable convolutions to save compute
View answer
✓ C. Stacking small 3×3 convolutions for a big receptive field
Two 3×3 layers see a 5×5 area with fewer weights and an extra non-linearity.
Question 5 of 9
You have only a few hundred labelled images. What's the usual approach?
- A. Start pretrained, freeze most layers, fine-tune the head
- B. Train a CNN from scratch on the few images
- C. Collect no more data and rely only on data augmentation
- D. Use a random forest on the raw pixels
View answer
✓ A. Start pretrained, freeze most layers, fine-tune the head
Pretrained features transfer well, so small datasets can still reach good accuracy.
Question 6 of 9
How does a Vision Transformer (ViT) process an image?
- A. It converts the image into a text caption first
- B. It reads the image pixel by pixel with an RNN
- C. It splits it into patches and runs a Transformer on them
- D. It uses only convolutions, with no attention layers
View answer
✓ C. It splits it into patches and runs a Transformer on them
With less built-in image bias than a CNN, ViTs usually need more data or pretraining.
Lesson: Vision Transformers
Question 7 of 9
How is an object detector usually evaluated?
- A. mAP, counting boxes right above an IoU threshold
- B. RMSE between predicted and true box sizes
- C. The number of boxes it finds in each image
- D. Accuracy of the predicted class for each box
View answer
✓ A. mAP, counting boxes right above an IoU threshold
IoU decides whether a box matches; mAP summarises precision and recall across classes.
Lesson: Object Detection
Question 8 of 9
Which architecture is a classic for image segmentation?
- A. U-Net, with skip connections for fine detail
- B. Word2Vec, trained on image captions
- C. AlexNet, the first deep network to win ImageNet
- D. A plain multi-layer perceptron on pixels
View answer
✓ A. U-Net, with skip connections for fine detail
The decoder upsamples to full resolution, and the skips bring back the edges lost by downsampling.
Lesson: Image Segmentation
Question 9 of 9
How can a tiny, invisible change to an image fool a classifier?
- A. It removes colour information from the image
- B. It changes the image size that the model sees
- C. Tiny shifts towards higher loss add up across pixels
- D. It can't: classifiers simply ignore tiny changes
View answer
✓ C. Tiny shifts towards higher loss add up across pixels
Methods like FGSM use the gradient to find these adversarial examples.