A channel is one parallel plane of values in an image or neural-network representation. A grayscale image has one channel; an RGB image has separate red, green, and blue channels stacked together. section 12.8 section 12.14
Channels keep different kinds of information separate while preserving the same pixel layout. Each RGB channel can be normalized independently, and CNN layers can combine multiple input channels into new feature maps. section 12.8 section 12.17
An RGB pixel across channels
Suppose one pixel has red 100, green 150, and blue 250. Its three channel values normalize to approximately -0.22, 0.18, and 0.96. Repeating this for every pixel produces three normalized matrices. section 12.8
Channel counts describe what a tensor contains:
- An 8×8 grayscale image is represented with one channel. section 12.14
- An 8×8 color image is represented with three channels: red, green, and blue. section 12.14
- A PNG image with transparency can use a fourth alpha channel. section 12.14
- A later convolution can process eight input channels together, summing their separate convolution results into each output value. section 12.17
Visualizes how convolutional layers produce multiple feature maps, making channels concrete as parallel learned representations. section 12.14
