What is the role of 1x1 convolution in CNN architectures?

1x1 convolutions are used for channel-wise feature transformation, dimensionality reduction, and increasing non-linearity.

A 1x1 convolution operates across channels without affecting spatial dimensions. It allows CNNs to mix information across feature channels efficiently. It is used in architectures like InceptionNet to reduce computational cost and increase depth without increasing spatial complexity. It also introduces additional non-linearity when followed by activation functions.