# Convolutional Neural Networks (CNN)
Confidence: high
Last verified: 2026-05-22
Generation: human_only


## TL;DR

CNNs are specialized neural networks for grid-structured data (images, video). Key operations: convolution (slide filters over input, detect patterns), pooling (downsample, reduce parameters), fully connected layers (classification). AlexNet (2012) won ImageNet by 10%+ margin, launching the deep learning revolution.

## Core Explanation

Convolution: filter/kernel learns spatial hierarchies — early layers detect edges, later layers detect objects. Pooling: max pooling (take max in window), average pooling. Key architectures: VGG (deep, simple), ResNet (skip connections, enable 152+ layers), Inception (multi-scale), EfficientNet (neural architecture search). Transfer learning: use pre-trained CNN, fine-tune on new task.

## Further Reading

- [ImageNet Classification with Deep Convolutional Networks (AlexNet, Krizhevsky et al., 2012)](https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks)