Channels, responses and feature maps AI
A convolution applies a learned local computation across an image or an incoming set of feature maps. An output channel is one resulting feature. Its values across spatial positions form a feature map; a response is the value at one of those positions.
A filter normally combines positions across incoming channels, rather than processing only one input color. Weight sharing applies that same filter at different locations. Stride and later downsampling reduce spatial resolution. See the convolution definition.
For a 224 × 224 crop in this report, layer4 produces 512 maps of 7 × 7 cells. Selecting channel 12 shows one 49-value map, not a complete image representation. Those cells correspond to broad receptive fields and are not an exact segmentation of the original pixels.
The exported maps are display images with fixed per-channel scales across the saved sample. Compare a channel across images; brightness across different channels is not directly comparable. No channel is automatically an “ear detector” or “cat neuron.” Such interpretations require further evidence.