Convolutional Neural Network
Subject: Computer and Technology noun /ˌkɑːn.vəˈluː.ʃən.əl ˈnʊr.əl ˈnɛtˌwɜːrk/ 1 definition
- deep-learning
- neural-networks
- computer-vision
- machine-learning
-
Definition (noun)
Entry #288 · UpdatedA Convolutional Neural Network (CNN) is a specialized deep learning architecture designed to process grid-structured data—such as 1D audio waveforms, 2D images, and 3D volumetric scans—by exploiting local spatial correlations. Rather than connecting every input unit to every hidden neuron, convolutional layers slide small learnable filter kernels across the input tensor using shared weights, computing local dot products followed by nonlinear activations such as ReLU to produce multi-channel feature maps. Successive stages of convolution and spatial downsampling (via pooling layers or strided kernels) build local translation invariance and compose low-level patterns like edges and textures into high-level semantic representations for downstream classification, detection, or segmentation. By reusing the same filter weights across every spatial position, CNNs require far fewer parameters than dense multilayer perceptrons on high-dimensional grids, though their fixed local receptive fields limit long-range global context modeling unless extended with deep residual stacks, dilated kernels, or attention layers.
Common Types
Residual CNN: Introduces identity skip connections (
y = F(x) + x) across convolutional blocks so gradients propagate cleanly through dozens or hundreds of layers without optimization degradation.Depthwise Separable CNN: Factorizes joint spatial-channel filtering into per-channel depthwise convolutions followed by
1 × 1pointwise projections, sharply cutting compute and memory footprint for mobile and edge inference.Fully Convolutional Network (FCN): Replaces fixed-size dense classification heads with
1 × 1convolutions and upsampling decoders (such as U-Net) to accept arbitrary input dimensions and emit pixel-level segmentation maps.1D and 3D Volumetric CNN: Adapts convolutional kernels to one-dimensional sequential signals (raw audio, ECG traces, or sensor streams) or three-dimensional grids (CT/MRI scans and spatio-temporal video), preserving locality along each axis.
Example
An automated electronics inspection pipeline feeds
640 × 640 × 3RGB circuit-board images into a PyTorch ResNet-34 CNN whose initial7 × 7stride-2 convolution produces 64 feature maps before four residual stages downsample the activation grid to20 × 20 × 512. Global average pooling collapses those spatial maps into a 512-element embedding vector, allowing a final linear head to classify solder-bridge defects in under8 msper frame on an edge GPU.