Feature Map

A feature map is the grid of values a convolutional layer produces after scanning its learned filter across an input. You can think of it as the layer’s report on where one particular learned pattern appears and how strongly it appears at each location.

How it is produced
A convolutional filter is a small set of learned weights, such as a 3×3 window. It slides across the input and, at every position, multiplies corresponding values and adds the results. One filter produces one two-dimensional feature map; a layer with 64 filters produces 64 maps, usually stored as a tensor shaped (height, width, channels) or (channels, height, width) depending on the framework. Early-layer maps can respond to simple local patterns, while deeper maps combine prior maps into more abstract patterns.

What “feature” means here
The word does not mean that a map has a guaranteed human-readable label such as “edge” or “corner.” A feature is simply a pattern useful for reducing the training loss. In a CNN, the output is commonly passed through an activation such as ReLU, so the map records positive evidence for that filter’s pattern. Its spatial arrangement is preserved: a high value near the upper left says the relevant pattern was detected near the upper left of the input.

Why feature maps shape a network
Their size and number control both what the network can retain and its computational cost.

Stride and pooling shrink feature maps, reducing memory and compute but discarding precise location detail.
Padding helps preserve map size and lets filters inspect border positions fairly.
More channels create more feature maps, increasing representational capacity but also parameters and activation memory.
In a ResNet block, skip connections carry feature maps forward, helping useful information and gradients survive depth.

A map that becomes too small too early can leave later layers unable to distinguish nearby structures. Conversely, very large maps across many channels can exhaust GPU memory. In PyTorch, these are the activations returned by layers such as torch.nn.Conv2d; during training, they must also be retained or recomputed for backpropagation.

A feature map is the array of activations produced when a convolutional filter is applied across an input. Each map records where and how strongly that filter detects its learned pattern; a convolutional layer outputs one map per output channel. Feature maps are the intermediate representations a network transforms into progressively more abstract features, and their size and number determine both receptive-field coverage and computational cost.

Imagine looking at a photo through several different colored highlighters. One might make edges stand out, another might reveal textures, and another might draw attention to round shapes. A feature map is like one of those highlighted versions of an image inside an AI system.

It is a grid showing where a particular visual clue appears and how strongly it appears. Early feature maps may highlight simple things such as lines or color changes. Later ones can highlight more meaningful patterns, such as eyes, wheels, or fur. By creating many feature maps, the network builds a richer picture of what is in an image.