feature channels

Feature channels are the depth dimension of a feature volume — the stack of feature maps at a given layer. If height and width are the 'where', channels are the 'what': each channel is the output of one filter, so it is a different learned feature detector. A tensor of shape (C, H, W) is best read as C separate H-by-W images, each lighting up where its own feature occurs. At the input, channels are the raw color planes (3 for RGB); inside the network, they are abstract learned features.

Precisely, the number of channels is a width hyperparameter set per layer. A convolutional layer takes C_in input channels and produces C_out output channels, where each output channel is computed by one filter that reads all C_in input channels at once. Channel counts typically grow with depth (e.g. 64, 128, 256, 512 in ResNet) as the network trades spatial resolution (shrinking H, W via stride/pooling) for representational richness (more, more-abstract features) — a pyramid that keeps the per-layer compute roughly balanced.

Channels are not independent silos: convolution explicitly mixes them, since every filter sums across all input channels, and 1x1 convolutions exist precisely to recombine channels cheaply. This channel mixing is where a network composes simple features into complex ones (combine an edge channel and a color channel into a 'red boundary' channel). Modern architectures manage channels deliberately — bottleneck blocks squeeze channels down then expand them, squeeze-and-excitation reweights channels by global importance, and grouped/depthwise convolutions restrict which channels talk to which to save computation.