Convolutional Layers

Question 1

A convolutional layer applies…

Accepted Answer

Small learnable filters — Each filter convolves over the image, computing dot products with local patches to detect a pattern, producing a feature map.

Question 2

WEIGHT SHARING in CNNs means…

Accepted Answer

The same filter weights reused everywhere — A filter's weights are shared across the whole image, drastically reducing parameters and detecting a pattern wherever it appears.

Question 3

The 'local receptive field' refers to…

Accepted Answer

The small region a filter looks at — Each convolution output depends only on a local patch, capturing local structure (edges, corners) efficiently.

Question 4

Why are CNNs more parameter-EFFICIENT than dense networks for images?

Accepted Answer

Weight sharing reuses one filter across positions — A dense layer on a 224×224 image needs millions of weights per neuron; a small shared filter needs only k×k per channel.

Question 5

CNNs provide approximate TRANSLATION INVARIANCE because…

Accepted Answer

A shared filter detects patterns anywhere — Since the same filter scans the whole image, a feature is detected at any location — useful since objects can appear anywhere.

Question 6

STRIDE controls…

Accepted Answer

How many pixels the filter advances per step — A stride of 2 moves the filter two pixels at a time, downsampling the feature map by roughly half.

Question 7

PADDING (e.g. 'same' padding) is used to…

Accepted Answer

Add border pixels to preserve the spatial dimensions — Zero-padding the borders prevents the feature map from shrinking each layer and lets the filter cover edge pixels fully.

Question 8

Each filter in a conv layer produces…

Accepted Answer

One feature map — Applying N filters yields N feature maps; deeper layers stack many to represent richer, more abstract features.

Question 9

The filters in a CNN are…

Accepted Answer

Learned from data via backpropagation — Unlike classical hand-crafted filters (e.g. Sobel), CNN filters are trained, discovering whatever patterns help the task.

Question 10

Early convolutional layers typically learn to detect…

Accepted Answer

Low-level edges and color gradients — A feature hierarchy emerges: early layers detect edges/textures, deeper layers combine them into parts and objects.

Convolutional Layer Visualizer

Input Image (5×5)

Kernel/Filter (3×3)

Output Map (3×3)

Slide one kernel

Test your knowledge

A convolutional layer applies…

Learn these first

Interview Questions

References & Further Reading

Love this visualization?