The architectures behind every AI model
Transformer, CNN, RNN, GAN, VAE, ResNet — every neural network architecture explained with what it's good at and where it fits in modern AI.
Deep learning architectures are the neural network "shapes" — how layers connect, what information flows where, how the network processes different data types. The Transformer dominates modern AI, but understanding it in context requires knowing the architectures it replaced (RNNs, LSTMs), the architectures that share ideas with it (CNNs, GNNs), and the specialized architectures that still matter (diffusion U-Nets, GANs).
These 60 terms are grouped into four sub-topics: sequence architectures (Transformer, RNN, LSTM, and modern successors), vision architectures (CNNs, ViTs, U-Nets), generative architectures (GANs, VAEs, diffusion), and cross-cutting concepts (attention variants, positional encodings, normalization). For newer alternatives to the transformer, see modern architectures.
The most important terms in Deep Learning Architectures
Start here if you're new. These entries explain the foundational vocabulary in depth.
Transformer
The dominant architecture in modern AI. Built around self-attention, it processes sequences in parallel rather than one step at a time — enabling the training of massive LLMs.
Read the full entry →VisionConvolutional Neural Network (CNN)
Neural network architecture that uses learnable filters (convolutions) to process grid-like data. Dominated computer vision for a decade and still used in image, video, and audio processing.
Read the full entry →SequenceRecurrent Neural Network (RNN)
Architecture for sequences that maintains a hidden state across time steps. Superseded by transformers for most tasks but still relevant for some streaming and low-resource cases.
Read the full entry →SequenceLong Short-Term Memory (LSTM)
A specific RNN variant with gates that help learn long-range dependencies. Dominant for sequences from 1997 until transformers took over around 2018.
Read the full entry →GenerativeGenerative Adversarial Network (GAN)
Two networks competing — a generator trying to create realistic samples and a discriminator trying to tell fake from real. Powered a decade of image synthesis before diffusion took over.
Read the full entry →VisionU-Net
A CNN architecture with encoder-decoder structure and skip connections. Originally for biomedical segmentation, now the backbone of every diffusion image model.
Read the full entry →All Deep Learning Architectures terms, organized
Grouped into sub-topics so you can find neighbors and prerequisites, not just alphabetical entries.
Sequence architectures
Vision architectures
Generative architectures
Cross-cutting concepts
Related term categories
These categories connect naturally to Deep Learning Architectures — many terms cross-reference between them.
Go deeper on Deep Learning Architectures
📖 Related concept tutorials
Long-form guides that walk through how these concepts actually work.
- How Transformers Work
The most important architecture, walked through.
- How Diffusion Models Work
The math and intuition behind image generation.
- How CNNs Work
Convolutions, pooling, and vision architectures.
🎯 Related models & tools
Real products and models where you'll encounter these terms.
- Claude Opus 4.8
Transformer-based frontier LLM.
- GPT-5
Transformer-based frontier LLM.
- Stable Diffusion
Latent diffusion image generation.
Questions about Deep Learning Architectures
Nearly, but not entirely. Diffusion U-Nets dominate image generation, CNNs still appear in vision preprocessing, and newer alternatives (Mamba, State Space Models) are gaining ground. But transformers dominate language, and hybrid architectures increasingly combine them with others.
Not first — start with transformers. But understanding RNN limitations (sequential processing, vanishing gradients) is what motivates the transformer's design decisions.
Encoder-only (BERT-style) is best for understanding tasks. Decoder-only (GPT-style) is best for generation and is what all modern chat LLMs use. Encoder-decoder (T5-style) is used for translation, summarization, and structured transformations.
Most current frontier ones, yes. GANs still power some real-time applications and video work, but the quality and controllability of diffusion have made it dominant.