Summary and Revision Notes

Key ideas

• Convolutional layers use local connectivity.
• A kernel is a small set of weights reused across positions.
• A feature map records detector responses.
• Padding changes border handling.
• Stride changes how far the kernel moves.
• Channels are parallel grids of values.
• A receptive field is the input region that can affect a feature.
• Parameter sharing reduces parameter count and reuses detectors.
• Pooling summarizes local regions.
• Convolution fits grids when local neighborhoods matter.

Common formulas

valid_output_length = floor((input_length - kernel_size) / stride) + 1
kernel_parameters = kernel_height * kernel_width * input_channels
patch_size = height * width

Common mistakes

• Treating CNNs as only a list of famous vision architectures.
• Forgetting that the same kernel weights are reused.
• Confusing input channels with output feature maps.
• Assuming convolution fits any table of numbers.
• Treating receptive field as a guarantee of understanding.

Before moving on

You should be able to compute a small convolution output, count kernel parameters, explain padding and stride, and describe why parameter sharing matters.