Notes
🔍

Channel

A channel is one slice of information carried at every spatial location in a convolutional network. An RGB image begins with three channels—red, green, and blue—but inside a network, channels become learned feature streams: one can respond to edges, another to textures, another to more abstract patterns.

How channels work in a convolution
A convolutional layer receives a tensor shaped roughly (batch, channels, height, width). Its filters look across all input channels, not just one. For example, a layer with 3 input channels, 64 output channels, and 3×3 kernels has weights shaped (64, 3, 3, 3). Each of the 64 filters combines a 3×3 patch from all three input channels, sums the result, and produces one output channel—also called a feature map.
The next layer treats those 64 feature maps as its input channels. This is how a CNN progressively recombines simple signals into richer representations.

Channels are capacity and communication
Increasing the number of channels gives a layer more distinct features it can represent. It also raises cost: a standard convolution’s parameter count grows with both input and output channels. A 3×3 layer going from 128 to 256 channels has 294,912 weights before biases. Channels therefore consume memory for activations and compute during training.
Useful channel-related designs include:

  • 1×1 convolutions, which mix information across channels at each location while leaving spatial size unchanged.
  • Depthwise convolutions, which process each input channel separately, followed by a pointwise 1×1 convolution to mix them; this greatly reduces compute.
  • BatchNorm, which commonly keeps separate learned scale and offset values per channel.

Practical consequences
In PyTorch, nn.Conv2d(in_channels, out_channels, ...) makes this choice explicit. A mismatch—feeding a tensor with 64 channels into a layer expecting 32—causes an immediate shape error. More subtly, making every layer very narrow can create a representation bottleneck: the network has too few feature streams to preserve useful signals. Making it excessively wide can slow training and overfit. Architectures such as ResNet carefully choose channel widths, and use 1×1 projections in skip connections when the number of channels must change.

In a convolutional network, a channel is one dimension of a tensor that holds a distinct feature representation at each spatial location. Input channels provide separate signals; output channels correspond to filters and form separate feature maps. A convolution combines all input channels for each output channel. Channel count determines a layer’s representational capacity, parameter count, and computational cost.

Think of an image as several transparent sheets stacked together. A normal color photo usually has three sheets, or channels: one showing how much red, green, and blue is present at every spot. Together, those sheets create the full-color picture.

In a convolutional network, channels can also be invisible “views” of an image. Early channels may highlight edges, bright areas, or textures. Later ones can capture more meaningful patterns, such as eyes, wheels, or fur. Each channel lets the network keep track of one kind of useful clue, and combining many channels helps it recognize what it is seeing.