It combines the channels into a single channel.
It discards all but the first channel.
It applies pooling independently to each channel.
It trains separate weights for each channel.
Fig: 1