✦ScienexySCIENCE KNOWLEDGE LIBRARY
Science reference · SCI-CS-003-111

Convolutional Neural Networks (CNNs)

Convolutional Neural Networks are deep learning architectures designed to process grid-like data, especially images, by learning spatial patterns and hierarchical visual features.

Y(i,j) = Σ_m Σ_n X(i+m,j+n)K(m,n) + b

Explanation

A Convolutional Neural Network (CNN) is a neural network architecture commonly used for processing images and other spatially structured data. Unlike a fully connected network, a CNN does not connect every input value to every neuron in the next layer. Instead, it applies small learnable filters, called kernels, across an input to identify local patterns.

The convolutional layer produces feature maps that indicate where particular patterns appear. Early layers may detect simple features such as edges and corners, while deeper layers can combine these features into more complex representations such as textures, shapes, and object parts.

CNN architectures often include several important components:

1. Convolutional layers: Apply learnable filters to extract spatial features.
2. Activation functions: Introduce nonlinearity, commonly using ReLU.
3. Pooling layers: Reduce spatial dimensions and computational requirements.
4. Normalization layers: Help stabilize and accelerate training.
5. Fully connected or alternative output layers: Convert extracted features into predictions.

Important convolution parameters include kernel size, stride, padding, and the number of filters. These determine how filters move across the input and the dimensions of the resulting feature maps.

CNNs are effective because the same filter is reused across different image locations. This weight sharing reduces the number of parameters compared with fully connected architectures and allows a network to recognize similar patterns in different positions. However, standard CNNs are not automatically invariant to every rotation, scale, or transformation.

Variables

SymbolMeaning

X

Input image or feature map

K

Learnable convolution kernel or filter

Y

Output feature map

b

Bias term

i,j

Output spatial coordinates

m,n

Kernel coordinates

Example or application

A CNN trained on animal photographs can learn to detect edges, fur textures, ears, and other visual patterns before combining them to classify an image as a cat, dog, or another animal.

Text

CNNs use convolutional filters to detect features such as edges, textures, shapes, and objects. Their local connectivity and shared weights make them particularly effective for image-related tasks.

Key Points

  • CNNs are especially useful for image and spatial data.

  • Convolutional filters learn local patterns.

  • Feature maps represent detected features.

  • Weight sharing reduces the number of parameters.

  • Pooling can reduce spatial dimensions.

  • Deeper layers learn increasingly complex representations.

Applications

  • Image classification

  • Object detection

  • Facial recognition

  • Medical image analysis

  • Autonomous vehicle perception

  • Satellite image analysis

  • Industrial visual inspection

Learning Outcomes

  • Explain the purpose of convolutional layers.

  • Identify kernels, feature maps, and pooling layers.

  • Describe how CNNs learn hierarchical visual features.

  • Understand the role of weight sharing.

Practice Questions

  • What is the main purpose of a convolutional filter?

  • How does weight sharing reduce model complexity?

  • What is the difference between convolution and pooling?

  • Why are CNNs effective for image classification?

Related Concepts

  • Artificial Neural Networks

  • Feature Extraction

  • Activation Functions

  • Deep Learning

  • Image Classification

Common Mistakes

  • Assuming convolution filters are manually designed rather than learned during training.

  • Confusing pooling with convolution.

  • Assuming CNNs are automatically invariant to all image transformations.

Sources and verification

Sources are shown as provided in the project data. Not every source has yet been independently verified. Please report errors or missing citations.

  • Goodfellow, Bengio, and Courville, Deep Learning, MIT Press, 2016.
  • LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P., Gradient-Based Learning Applied to Document Recognition, Proceedings of the IEEE, 1998.

Suggest a correction or source · More in Artificial Intelligence & Machine Learning