Week 8: Convolutional Neural Networks — Knowledge Check

Question 1

A) CNNs use a lower-precision floating point representation for their weights.
B) CNNs use local connectivity and weight sharing — the same filter weights are applied across all spatial positions in the input.
C) CNNs skip connections between non-adjacent layers to reduce the total parameter count.
D) CNNs process only the center region of the image and ignore the borders.

Question 2

A) The output feature map doubles in spatial size because the filter covers more of the input.
B) The output feature map spatial dimensions are roughly halved because the filter steps over the input less frequently.
C) The number of learnable parameters in the layer doubles because more positions are computed.
D) The receptive field of each output neuron decreases proportionally to the stride.

Question 3

A) It improves model accuracy by giving the network additional information about the input's boundaries.
B) It allows the output feature map to retain the same spatial dimensions as the input and prevents information loss at the borders.
C) It reduces the receptive field of filters so each neuron focuses on a smaller region.
D) It prevents gradient explosion during backpropagation through convolutional layers.

Question 4

A) Max pooling forces the network to learn rotationally symmetric filters.
B) By selecting the maximum activation within a local region, max pooling makes the detection of a feature robust to small shifts in its exact position.
C) Max pooling normalizes feature map values so that spatial location does not affect the magnitude of activations.
D) Max pooling adds trainable parameters that learn to compensate for positional shifts in the input.

Question 5

A) ReLU produces bounded outputs in the range [0, 1], which prevents gradient explosion.
B) ReLU does not saturate for positive inputs, so gradients flow without diminishing through many layers, while sigmoid and tanh saturate and cause vanishing gradients.
C) ReLU is differentiable everywhere, making backpropagation mathematically simpler than with sigmoid.
D) ReLU reduces the number of parameters in a layer by zeroing out a fixed fraction of weights during training.

Question 6

A) Deeper networks require more training data and were overfitting to what they had
B) The optimization process itself was struggling — adding more layers made the network harder to train, not just harder to generalize
C) Deeper networks are slower to train and ran out of compute budget before converging
D) The activation functions stopped working correctly beyond a certain network depth

Question 7

A) Small adjustments require fewer parameters, making the model more compact
B) If no adjustment is needed, the block can simply output zero — learning to do nothing is easy, whereas learning an identity mapping through nonlinear layers is surprisingly hard
C) Small adjustments are computed faster because they use fewer matrix operations
D) Skip connections enforce that adjacent blocks learn orthogonal, non-redundant features

Question 8

A) Training only the last few convolutional layers while freezing the rest; appropriate when the target domain is moderately different from the source domain.
B) Freezing all pre-trained layers and training only a new linear classifier on top of the extracted features; most appropriate when the target dataset is very small and overfitting is a major risk.
C) Using a linear activation function in the final layer to produce regression outputs rather than classification probabilities.
D) Probing individual neurons in the network to determine which filters are responsible for specific visual concepts before deciding which layers to freeze.

Question 9

A) Freeze the final classification layer (since it is domain-specific) and fine-tune the earlier convolutional layers (since they contain general features).
B) Freeze the early convolutional layers (which learn general low-level features like edges and textures that transfer well) and fine-tune the later layers (which learn high-level, domain-specific features).
C) Fine-tune all layers equally with the same learning rate to ensure the model adapts to the new domain as quickly as possible.
D) Freeze all layers and only train the batch normalization statistics, which is sufficient to adapt to a new domain.

Question 10

A) Deeper layers detect simpler, more localized features such as single edges, while earlier layers detect complex object parts.
B) Earlier layers detect low-level features (edges, colors, simple textures) while deeper layers combine these into increasingly abstract and task-specific representations (object parts, entire objects).
C) All layers in a CNN detect the same types of features at different spatial scales simultaneously.
D) Earlier layers detect global context while deeper layers refine local texture details to sharpen the representation.