CNN Architecture (Full Pipeline)

Question 1

What is the purpose of the Flatten layer in a CNN?

Accepted Answer

Convert 2D feature maps into a 1D vector — The Flatten layer reshapes the 2D feature maps into a 1D vector, which can then be fed into fully connected (dense) layers for classification.

Question 2

Why do we typically increase the number of filters in deeper convolutional layers?

Accepted Answer

To capture more complex, abstract features — Deeper layers capture more complex, abstract features. More filters allow the network to learn a richer variety of these high-level patterns.

Question 3

A classic CNN architecture is built by stacking repeated blocks of…

Accepted Answer

Convolution, activation, and pooling blocks, repeated — The canonical pattern alternates convolution (feature extraction), non-linearity, and pooling (downsampling), repeated to build depth.

Question 4

Through a CNN, spatial dimensions generally ___ while channel depth ___.

Accepted Answer

shrink; grows — Pooling/striding reduce height×width while added filters increase depth — trading resolution for richer semantic features.

Question 5

The feature HIERARCHY a CNN learns goes from…

Accepted Answer

Low-level edges toward abstract objects — Early layers detect simple patterns; deeper layers compose them into increasingly abstract, task-relevant concepts.

Question 6

After the convolutional/pooling stack, a classifier 'head' typically…

Accepted Answer

Flattens — The extracted feature volume is reduced to a vector and passed to fully connected layers ending in softmax for class probabilities.

Question 7

AlexNet (2012) was historically pivotal because it…

Accepted Answer

Won ImageNet decisively using deep GPU-trained CNNs — AlexNet's GPU-trained deep CNN dramatically beat classical methods on ImageNet, igniting modern deep learning in vision.

Question 8

VGG networks are known for…

Accepted Answer

Using small 3×3 kernels stacked deeply — VGG showed that stacking many small 3×3 convolutions builds large receptive fields with fewer parameters than big kernels.

Question 9

ResNet's key innovation is the…

Accepted Answer

Skip connections (residual blocks) — Skip connections let gradients/identity bypass blocks, solving degradation and allowing hundreds of layers to train successfully.

Question 10

Why do deeper CNNs (with skip connections) tend to outperform shallow ones?

Accepted Answer

They learn a deeper hierarchy of abstract features — Depth enables richer compositional features; residual connections make that depth trainable without vanishing gradients.

Animation

Architecture

Network Stats

A classic CNN, end to end

Test your knowledge

What is the purpose of the Flatten layer in a CNN?

Learn these first

Interview Questions

References & Further Reading

Love this visualization?