Convolutional Neural Networks
Use the experiment to inspect how filters, stride, padding, pooling, and receptive field transform an image through a convolutional network.
What to change
Change the kernel, stride, padding, and pooling controls one at a time so each visual change has a clear cause.
What to inspect
Follow spatial dimensions, activation maps, border effects, and the effective receptive field through the network.
What to conclude
A CNN design is a sequence of information-preserving and information-compressing choices, not just a stack of layers.
CNN FAQ
What does a convolutional filter learn?
A filter learns a local pattern that can be detected across spatial locations, such as an edge, texture, or higher-level feature.
Why is weight sharing useful?
The same filter is applied across the image, reducing parameters and making detection less dependent on absolute position.
What does stride change?
Stride controls how far the filter moves. Larger stride reduces spatial resolution and computation but may discard detail.
What does padding change?
Padding controls how borders are treated and whether output dimensions shrink after convolution.
What is a receptive field?
It is the region of the original input that can influence a particular activation deeper in the network.
Why use pooling?
Pooling summarizes local activations and reduces spatial size, but excessive pooling can remove precise location information.
Are larger kernels always better?
No. Larger kernels see more context but cost more parameters and may be replaced by multiple smaller convolutions.
How do I compare CNN architectures fairly?
Keep data, preprocessing, evaluation splits, and training budgets fixed, then compare accuracy, calibration, latency, and robustness.