Convolutional Neural Network

A neural network that shares small learnable filters across an input to detect local patterns efficiently.

Definition

A convolutional neural network (CNN) applies small learnable filters that slide across the input, detecting local patterns such as edges or textures regardless of position. Stacking convolution and pooling layers builds up from simple features to complex ones.

Deeper layers of a trained CNN respond to increasingly abstract patterns, from edges to textures to object parts, a hierarchy that can be visualized and that mirrors ideas from biological vision. This learned feature hierarchy is what replaced decades of hand-designed image features.

The inductive bias toward local, translation-invariant patterns is what lets CNNs learn from far less data than a fully connected network would need on images. That same bias limits them when spatial position carries meaning that should not be ignored, or when relationships span the whole input; hybrid designs and attention address these cases, but the convolution remains the efficient workhorse for grid-structured data.

Key ideas

• Weight sharing: one filter reused across all positions, cutting parameters.
• Local receptive fields: each unit sees a small region.
• Pooling: downsamples to add translation tolerance.
• Translation equivariance: a shifted input gives a shifted response.

Why it matters

CNNs made large-scale image recognition practical and remain central to vision, medical imaging, and any data with grid structure. Their inductive bias toward local, shift-invariant patterns is what lets them learn from images efficiently.

Fusion connection

CNNs analyze spatial fields from plasma simulations, such as magnetic equilibrium or heat-flux maps, extracting features that correlate with Hyperion performance and stability.