Opening the paper…

Deep Learning for Computer Vision, Quiz 2

January 2026 term, 12 Apr 2026, Set 1

Learning mode

Question 1

+1−0Single correct

Why do very deep plain CNNs (without skip connections) sometimes show higher training error than shallower CNNs?

Select one option

A.
The increased capacity leads to severe overfitting on the training set before convergence can be reached.
B.
Optimization becomes harder; skip connections make learning identity mappings easier and improve gradient flow.
C.
The receptive field becomes excessively large, causing the network to lose fine-grained spatial information.
D.
Saturated activations cause exploding gradients that cannot be mitigated by standard initialization techniques.

Report a problem

to report a problem with this question.

Choose a question

Answered: 0
Not answered: 1
Marked for review: 0
Answered & marked for review: 0
Not visited: 22

Discussion

Submit this paper?

You cannot change your answers after submitting.

Answered: 0
Not answered: 1
Marked for review: 0
Answered & marked for review: 0
Not visited: 22

Instructions

The palette shows the status of every question. Pick a number to go straight to it.

Answered: 0
Not answered: 1
Marked for review: 0
Answered & marked for review: 0
Not visited: 22

Save & Next moves to the next question. Answers are saved as you go.
Mark for Review & Next flags the question to come back to. A flagged answer still counts.
Clear Response removes your answer to the question on screen.
The ← and → keys move to the previous and next question.

Why do very deep plain CNNs (without skip connections) sometimes show higher training error than shallower CNNs? In a standard ResNet-50 bottleneck block, the three convolutions are typically: In Inception/GoogLeNet modules, the main purpose of 1×1 convolutions is to: