If you want to find artificial-intelligence patents, the single most useful piece of metadata is the Cooperative Patent Classification (CPC) symbol, and the class that anchors the field is G06N. The CPC scheme gives G06N the title "Computing arrangements based on specific computational models." That phrasing is deliberately broad — it captures the family of computing approaches that depart from conventional sequential logic, including neural networks, probabilistic models, fuzzy logic, and quantum and other unconventional computing models. In practice, when a patent examiner classifies a machine-learning invention, G06N is where the primary symbol usually lands.

The CPC is a jointly administered classification system used by the USPTO and the European Patent Office, with roughly 250,000 hierarchical subdivisions. Each symbol narrows the subject matter: the letter (G) is the section, G06 is the class for computing and calculating, G06N is the subclass for specific computational models, and the numbers after the slash drill down into specific techniques. Understanding the G06N tree is the fastest way to navigate the AI patent landscape, because the subgroups map closely onto how practitioners actually think about the technology.

How G06N divides the field

The most heavily populated branch of G06N is G06N 3/00, which the CPC scheme titles "Computing arrangements based on biological models." Despite the biological framing, this is the neural-network branch — the subgroup that covers artificial neural networks of every kind, because such networks were originally conceived as models of biological neurons. Within it, two children do most of the work for AI filings. G06N 3/04 is titled "Architecture, e.g. interconnection topology" and covers the structure of a network: how layers and nodes are arranged and connected. G06N 3/08 is titled "Learning methods" and covers how a network is trained — backpropagation, gradient methods, and the procedures that adjust a model's weights.

That division — architecture versus learning method — is visible in real grants. A patent on a transformer's internal structure tends to carry G06N 3/04 (often the more specific 3/045 for neural-network architecture), while a patent on a training procedure carries G06N 3/08 (or 3/084 for backpropagation). A single AI patent frequently carries several G06N symbols at once because it claims both a structure and a way of training it.

"Computing arrangements based on specific computational models"— USPTO, CPC scheme G06N (class title), source

Alongside the neural-network branch sits G06N 20/00, titled "Machine learning." This is the general-purpose machine-learning subgroup, used for learning systems that are not specifically framed as biological/neural models — ensemble methods, supervised and unsupervised learning pipelines, and the like. The coexistence of G06N 3/00 and G06N 20/00 reflects a real distinction: a deep neural network is one family of machine learning, but machine learning as a discipline is broader, and the classification preserves that. Other G06N branches cover probabilistic and statistical models (G06N 7/00), knowledge-based and rule systems (G06N 5/00), and quantum computing (G06N 10/00), each of which catches a slice of AI-adjacent filings.

Why AI patents rarely live in G06N alone

Classification is rarely tidy for AI, because the inventions span more than the model itself. A machine-learning patent that claims a hardware accelerator — a chip designed to run inference efficiently — will typically carry a primary symbol in G06F (the subclass for digital data processing, including processor and memory architecture) with G06N as a secondary symbol, because the inventive contribution is in the hardware arrangement, not the model. Conversely, an AI patent whose contribution is in computer vision will pick up G06V (image or video recognition), with symbols such as G06V 10/82 marking the use of neural networks for image analysis. This is why a portfolio search restricted to G06N alone will systematically undercount AI activity: a large share of applied-AI inventions are dual-classified, and some are primarily classified in G06F or G06V with G06N riding along.

For anyone mapping an AI patent landscape, the practical implication is that the query matters as much as the count. Searching G06N captures the model-centric inventions; adding G06F catches the accelerator and inference-hardware layer; adding G06V catches the multimodal and vision layer. The CPC scheme is explicit that a single document can — and for AI frequently does — carry symbols across all three. A landscape that reports "how many AI patents company X holds" is only as meaningful as the CPC net it casts, and counts are sensitive to which subgroups are included and to filing dates, since classifications are assigned at publication.

A worked example makes the multi-symbol reality concrete. The Google grant US11790214B2, "Mixture of experts neural networks," carries two G06N symbols at once: G06N 3/045 (a neural-network architecture symbol within the architecture branch) and G06N 3/08 (learning methods). That dual marking is not redundancy — it reflects that the patent claims both a structural arrangement (an MoE subnetwork positioned between two layers) and the learned behavior that arrangement enables. A search that included only G06N 3/08 would still find it; a search restricted to G06N 20/00 would miss it entirely, because the invention is classified as a biological/neural model rather than under the general machine-learning symbol. This is the practical hazard of single-symbol searching: the same invention can be invisible or visible depending on which G06N child you query, which is why landscape work uses symbol sets rather than single codes and reports the query alongside the count.

The bottom line for the question is direct: the home class for AI patents is G06N, "Computing arrangements based on specific computational models," with neural networks in G06N 3/00, architecture in G06N 3/04, learning methods in G06N 3/08, and general machine learning in G06N 20/00. But the field bleeds into G06F for hardware and G06V for vision, and the most reliable AI searches treat those three subclasses together rather than relying on G06N in isolation. The CPC titles themselves, published and maintained by the USPTO and EPO, are the authoritative definitions of what each symbol covers.