Feature Map
A feature map in a convolutional neural network (CNN) is the output representation obtained after applying convolutional filters to the input data. It highlights patterns and structures such as edges, textures, or object parts, depending on the layer’s depth in the network.
Background and importance
Feature maps are central to how CNNs process data. Early layers focus on low-level features (edges, gradients), while deeper layers detect high-level concepts (faces, wheels, digits). By stacking feature maps across layers, CNNs learn hierarchical representations that make them highly effective for image and video analysis.
Use cases
- Image classification: cats vs dogs, handwritten digit recognition.
- Medical imaging: detecting tumors in CT or MRI scans.
- Explainable AI: feature map visualizations help developers verify whether the model focuses on the right parts of an image.
Pros and challenges
- Feature maps provide insights into what the network learns.
- However, with hundreds of maps per layer, interpretation can be overwhelming.
Feature maps can be seen as the “visual fingerprints” of data as it travels through a convolutional network. Each filter captures a specific aspect of the input—whether it’s a simple edge or a complex texture—and the resulting feature map highlights where and how strongly that pattern appears in the image.
One important property of feature maps is spatial locality: nearby pixels in the map correspond to nearby regions in the original image. This allows CNNs to preserve structural information, making them especially powerful for vision tasks. Pooling layers often follow convolutions, reducing the size of feature maps while keeping the most salient information.
In research and practice, feature map visualisation is widely used to interpret neural networks. By examining which areas “light up” in a feature map, practitioners gain insight into what the model attends to. This not only helps in debugging but also supports trust and accountability when AI is used in sensitive domains like healthcare or autonomous driving.
Further reading
- LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature.
