[discussion] Why increase channel size in CNN layers?

In most CNN architecture I see, they increase the channels from 1 or 3 to like 16, 32 or even 64. Isn’t the point to decrease the dimensionality?

architecture channels cnn dimensionality isn machinelearning