Principle:LaurentMazare Tch rs Convolutional Layer
| Knowledge Sources | |
|---|---|
| Domains | Computer_Vision, Neural_Network_Layers |
| Last Updated | 2026-02-08 14:00 GMT |
Overview
Layer that applies learned spatial filters over 2D input feature maps to extract local patterns through convolution operations.
Description
A 2D convolutional layer slides a set of learned kernels (filters) across the spatial dimensions (height, width) of an input tensor. Each kernel produces one output channel by computing element-wise multiplications and summing the results at each spatial position. Convolutions exploit spatial locality and translation equivariance, making them the foundation of image recognition architectures. Parameters include kernel size, stride, padding, dilation, and groups for depthwise or grouped convolutions.
Usage
Use this principle when processing 2D spatial data (images, feature maps) where local pattern detection is needed. Essential for image classification, object detection, and any task requiring spatial feature hierarchy extraction.
Theoretical Basis
Where denotes the 2D cross-correlation operation.
Parameters:
- Kernel: Weight tensor of shape [out_channels, in_channels/groups, kH, kW]
- Stride: Step size of the convolution
- Padding: Zero-padding added to input borders
- Dilation: Spacing between kernel elements