Feature Maps in Convolutional Neural Networks

The Strongest Local Evidence

A convolutional neural network does not need to preserve every spatial position forever. After earlier layers have detected useful patterns, the network can compress nearby positions while keeping the strongest evidence that a pattern is present. Max pooling performs this compression by examining small regions of a feature map and carrying forward the largest value from each region.

One Window One Output

A max-pooling window covers a local region of one channel of a feature map. The operation compares the values inside that region and writes the largest one into a single output position. The output therefore records the strongest local activation found by that window.

Window Movement and Size Reduction

The window size determines how much local area is examined at once. The stride determines how far the window advances before producing its next output. In the common 2 × 2, stride 2 setup, the window advances two positions at a time horizontally and vertically. This arrangement produces one output for each non-overlapping 2 × 2 region, reducing both spatial dimensions by a factor of 2.

Independent Channel Processing

A feature map can contain multiple channels, with each channel representing a different learned response. Max pooling processes each channel independently. The same local pooling rule is applied within one channel at a time, so values from different channels are not compared with one another. The spatial dimensions become smaller, but the channel dimension remains.

• Choosing one maximum across the entire feature map

  Max pooling makes many local decisions within each channel. It does not select one global maximum.

  Fix: Apply the window separately at each spatial location within each channel.

• Comparing values across channels

  Pooling operates independently by channel rather than selecting one value across the whole feature map.

  Fix: Preserve a separate pooled output map for every channel.

Cost and Spatial Hierarchy

Downsampling has two linked benefits. First, it reduces the number of feature-map coefficients that later layers must process. The source gives a 22 × 22 × 64 feature map containing 30,976 coefficients per sample. Flattening that map before a Dense layer of size 512 would create 15.8 million parameters, which would be far too large for the small model being discussed and could cause intense overfitting.

Second, repeated downsampling supports a spatial hierarchy. Later convolutional windows can cover an increasingly larger fraction of the original input. If feature maps remained large throughout the network, a 3 × 3 window in a later layer would still describe only a relatively small region of the original input. For tasks such as recognizing a whole digit, later features need access to information from across a larger portion of the input.

Three Ways to Reduce Space

Max pooling is useful when the strongest local activation is the important signal. The source describes feature maps as encoding the spatial presence of a pattern or concept across different tiles. Keeping the maximum emphasizes the strongest local presence. Average pooling instead combines the values into an average, which can dilute that evidence. A convolutional layer can also reduce spatial dimensions by using strides, but that reduction occurs through a convolutional layer rather than through a pooling summary.

Common Interpretation Errors

• Thinking max pooling preserves every value in a region

  The window is replaced by one value: the maximum.

  Fix: Track one output value for each local window.

• Assuming stride 2 means only the values are reduced

  The stride controls how far the window advances, and the common 2 × 2, stride 2 setup reduces each spatial dimension by a factor of 2.

  Fix: Track both the window size and its movement across height and width.

• Assuming pooling removes channels

  Pooling changes spatial size but operates independently for each channel.

  Fix: Keep one smaller spatial map for each original channel.

• Treating max pooling and average pooling as equivalent

  Max pooling keeps the strongest value, while average pooling can weaken or dilute strong evidence.

  Fix: Identify whether the operation selects a maximum or calculates an average.

Check Your Reasoning

Key Takeaways

1. Max pooling replaces each local window with its maximum value, preserving the strongest local activation.
2. A 2 × 2 window with stride 2 reduces height and width by a factor of 2.
3. Pooling operates independently within each channel, so it does not compare values across channels.
4. Downsampling reduces the number of coefficients later layers process and supports increasingly broad spatial features.
5. Average pooling summarizes a region with an average, while strided convolution reduces spatial dimensions through a convolutional layer.