A neural network architecture is a structured design specifying how computational layers are organized, interconnected, and oriented to process information. It defines the layer types (convolutional, recurrent, attention-based, fully connected), their arrangement (serial, parallel, residual, encoder-decoder), connectivity patterns, and the direction of data flow through the network. The architecture determines what transformations the network can represent and how efficiently it computes them — the architectural choices (depth, width, skip connections, attention heads) are the blueprint that distinguishes one class of neural model from another. Persistence mechanism: architectures are specified in technical papers, implemented in library code (PyTorch, TensorFlow), and reproduced across independent research and production systems, surviving the particular datasets and training runs they were built for.
[formal: architectura | substrate: mind | horizon: generations | explicit: yes | epoch: 0.01]