Architectures
Different problem shapes call for different model architectures. You will rarely design one, but knowing which family fits your data tells you whether a proposal is credible.
Neural Networks — the Base Concept
A neural network is layers of simple math units, each taking weighted inputs and passing a signal forward. Training adjusts those weights until the network's outputs match the examples it was shown. Everything below is a specialized arrangement of this same idea, shaped to fit a particular kind of data.
CNN — Images and Spatial Patterns
Convolutional Neural Networks scan an input with small filters that detect local patterns — edges, then textures, then objects. Because the same filter slides across the whole image, CNNs are efficient and translation-tolerant. They remain the workhorse for medical imaging, quality inspection, document scanning, and anything where position and shape carry meaning.
RNN, LSTM/GRU — Sequences and Time-Series
Recurrent networks process data one step at a time while carrying a memory of what came before, which suits sensor streams, demand forecasting, and log analysis. LSTM and GRU variants added gating to hold information over longer spans. They've been largely displaced by transformers for language, but remain practical and cheap for straightforward time-series work.
Transformer — Attention and the Modern LLM
The transformer replaced step-by-step recurrence with attention: every element of the input can directly weigh every other element, all in parallel. That parallelism is what made training on internet-scale data feasible, and the resulting models are the backbone of essentially every current large language model, plus much of modern vision and audio work.
GANs and Diffusion — Generating Images and Video
GANs pit a generator against a discriminator until the fakes are convincing. Diffusion models instead learn to reverse a noising process, starting from static and denoising toward a coherent image. Diffusion now dominates commercial image and video generation because it trains more stably and gives finer control over the result.
Match the architecture to the data shape: grids and images to CNNs, ordered sequences to recurrent or transformer models, language and multimodal work to transformers, synthetic media to diffusion.