Architectures
Neural network architectures encode assumptions about how information should be represented, exchanged, and refined. This section traces the major design families and the problems each was built to solve.
Pages are ordered by year of introduction, oldest first. Read chronologically to follow the evolution from general layered networks to recurrence, convolution, attention, specialized vision backbones, and modern generators.
- MLPMulti-Layer Perceptron
- RNNRecurrent Neural Network
- LSTM / GRULong Short-Term Memory and Gated Recurrent Unit
- CNNConvolutional Neural Network
- LeNet-5LeNet-5 - The pioneering CNN architecture
- AlexNetAlexNet - The CNN that reignited deep learning
- GANGAN - Generative Adversarial Networks
- GoogLeNet / InceptionGoogLeNet - Multi-scale feature extraction with Inception modules
- VGGNetVGGNet - Very deep CNNs with small 3×3 filters
- ResNetResNet - Deep residual networks with skip connections
- U-NetU-Net - Encoder-decoder architecture for image segmentation
- YOLOYOLO - Real-time object detection in a single pass
- DenseNetDenseNet - Densely connected convolutional networks
- MobileNetMobileNet - Efficient CNNs for mobile and edge devices
- TransformerSelf-attention based architecture
- EfficientNetEfficientNet - Compound scaling for efficient and accurate CNNs
- Diffusion ModelsGenerative models via denoising
- Vision Transformer (ViT)Transformer applied to image data
- PatchCoreIndustrial anomaly detection using a memory bank of normal patch features
- Swin TransformerSwin Transformer - Hierarchical vision transformer with shifted windows