CNNs & convolutional layers deep dive 15 questions 25 min

CNN & Convolutional Layers MCQ · test your knowledge

From filters and stride to modern architectures – 15 questions covering convolution arithmetic, pooling, receptive fields, and key CNN building blocks.

Easy: 5 Medium: 6 Hard: 4
Filters
Stride
Padding
Pooling

Convolutional Neural Networks: the eyes of deep learning

CNNs use convolutional layers to automatically learn spatial hierarchies of features. This MCQ tests your understanding of convolution operation, filter dimensions, pooling, and architectural patterns like residual connections.

Why convolutions?

Convolution preserves spatial structure, uses parameter sharing (translation equivariance), and is computationally efficient compared to fully connected layers. Perfect for images, video, and any grid‑like data.

CNN glossary – key concepts

Filter / Kernel

A small matrix (e.g., 3x3) that slides over the input to produce a feature map. Each filter detects a specific pattern.

Stride

Step size of the filter movement. Stride 2 downsamples the feature map (reduces width/height).

Padding

Adding zeros around the input to control output size. 'Same' padding preserves spatial dimensions.

Pooling (Max/Avg)

Downsamples feature maps, introduces translation invariance, and reduces computation.

Receptive field

The region of the input that influences a particular neuron. Deeper layers have larger receptive fields.

Feature map / channel

Output of one filter applied across the input. Multiple filters produce multiple channels.

1x1 convolution

Pointwise convolution, used to change channel depth and add non‑linearity without spatial mixing.

# Conv layer output size formula (PyTorch style)
# H_out = floor((H_in + 2*padding - dilation*(kernel-1) -1)/stride + 1)
# Classic example: 32x32 input, 5x5 filter, stride 1, padding 2 → 32x32 output (same padding)
Interview tip: Be ready to compute output dimensions after convolution/pooling, explain parameter sharing, and compare architectures like VGG vs ResNet. This MCQ covers these calculations and trade‑offs.

Common CNN interview questions

  • How do you calculate the number of parameters in a conv layer?
  • What is the effect of stride and padding on feature map size?
  • Why do we use pooling? Are there alternatives?
  • Explain the concept of receptive field. How can you increase it?
  • What are the benefits of 1x1 convolutions?
  • How do residual connections help train deeper networks?
AI Hub Next: RNN & LSTM