⟵ Back to Course Overview

Convolutional Layer Visualizer

Multi-Channel Feature Extraction

Task Description: Convolution operations are fundamental to understanding how CNNs learn. In this task, three filters with fixed parameters are used to extract different features from the input data. First filter or CNN kernel simply obtains the mean over all the values in a 3x3 region whereas the other two are a bit more sophisticated. CNNs learn the values of these kernels depending on the training data which gives them a lot of flexibility in solving complex tasks.

Input $\mathbf{X}$

$K_0$ (Mean)

$K_1$ (Edge)

$K_2$ (Sharpen)

$\mathbf{Y}_0$ (Blurred)

$\mathbf{Y}_1$ (Edges)

$\mathbf{Y}_2$ (Sharp)

The Mean Operation ($K_0$): Acts as a low-pass filter. Each output pixel is the average of its $3 \times 3$ neighborhood: $$ K_{0} = \frac{1}{9} \begin{bmatrix} 1 & 1 & 1 \\ 1 & 1 & 1 \\ 1 & 1 & 1 \end{bmatrix} $$

Padding Mode:

Instructions: Hover over any Output cell ($\mathbf{Y}$) to see the corresponding receptive field in the Input ($\mathbf{X}$).

The different kernels capture different spatial frequencies. $K_0$ smooths the signal, while $K_1$ and $K_2$ emphasize high-frequency changes.