\(1 \times 1\) convolutions operate on a single pixel without regard to any neighboring pixels. Thus, they can be used to reduce the number of channels (the 3rd dimension in \(H \times W \times C\) ).