Vision · 1989–2012 · foundational
Convolutional Neural Network
A network built from convolutional layers that exploit the spatial structure of images: local receptive fields, shared weights and hierarchical feature composition.
ModLensVision
Interactive Diagram
Focus the lens
Click any component to read what it does. Signal direction follows the edges.
inputImage. Raw pixels with explicit spatial structure — the assumption convolutions exploit.
Core idea
Translate invariance into architecture: a detector learned at one location applies everywhere, and deep stacks compose edges into textures, parts and objects.
Why it exists
Fully connected networks ignore spatial structure and explode in parameters; convolutions encode locality and translation sharing directly.
Mathematics
(f * k)(i,j) = \sum_m \sum_n f(m,n)\,k(i-m, j-n)
Data Flow
What moves through the system
- 01Pixels enter as a tensor with spatial dimensions.
- 02Convolutional layers extract local patterns with shared weights.
- 03Pooling or striding reduces resolution while increasing abstraction.
- 04A head maps the final representation to the task output.
Strengths
- + Parameter efficiency via weight sharing
- + Strong spatial inductive bias
- + Mature tooling and intuition
Limitations
- − Limited receptive field per layer
- − Long-range dependencies require many stacked layers
- − Uniform compute regardless of content
Applications
- · Classification
- · Detection backbones
- · Segmentation
- · Medical imaging
