Why do convolutional neural networks outperform fully connected networks in vision tasks?

Question

Options

• AThey reduce training time to zero
• BThey exploit local connectivity and weight sharing
• CThey remove the need for labeled data
• DThey guarantee interpretability

• A
  4% (1)
• B
  85% (23)
• C
  7% (2)
• D
  4% (1)

Explanation

CNNs are designed with two key inductive biases suited to image data: (1) Local connectivity - each neuron connects only to a small receptive field in the input, capturing spatially local features (edges, textures) rather than comparing every pixel to every other pixel; and (2) Weight sharing - the same filter (kernel) is applied across all spatial positions, dramatically reducing the number of parameters and allowing the network to detect the same feature regardless of its location (translation equivariance). Together these properties make CNNs far more parameter-efficient and effective than fully connected networks on high-dimensional image inputs. Options A, C, and D are all false: CNNs do not eliminate training time, still require labeled data, and are generally not more interpretable than fully connected networks.

No community discussion yet for this question.