Nvidia Patent Selects the Optimal Neural Network Architecture Through Federated Learning
Nvidia has filed a patent for a system that figures out which neural network design works best for a task, without any of the organizations involved ever having to share their raw data with each other.
How Nvidia trains AI across private data silos
Ever wondered how hospitals could contribute to a shared AI model without sending patient records outside their walls? That's exactly the problem this patent is trying to solve, and it turns out the usual answer, gathering all the data in one place and training there, simply isn't allowed in most regulated industries.
Nvidia's approach is to let each organization train its piece of the AI model on its own data, then share only the learned patterns (not the underlying records) back to a central coordinator. The coordinator uses those contributions to figure out not just how well the model is learning, but which model design works best across all the different data sets involved.
The clever part is that different organizations may be using different types of data, and the system can match which portions of the AI network are relevant for which data type. You end up with a better-tuned model without anyone ever handing over their private information.
… cause one or more parameters of different portions of one or more first neural networks to be updated using one or more parameters of correspondingly different portions of two or more second neural networks, wherein the different portions of the two or more second neural networks correspond to the type of training data …
Translation: The system updates parts of a neural network by pulling in specific data patterns from other networks trained on similar info.
How the federated architecture search actually selects a model
The patent describes a process called federated neural architecture search, which combines two existing ideas in a single workflow.
The first idea is federated learning (training a shared AI model across multiple independent computers or institutions, where only model updates travel over the network, never the raw data). The second is neural architecture search (automatically testing many possible AI network designs to find the one that performs best on a given task, rather than having engineers pick the structure by hand).
The system works roughly like this:
- Each participating organization (a hospital, a research lab, a company) trains a local version of the AI on its own private data.
- The type of data each site holds is noted, and only the portions of the model that are relevant to that data type are updated from that site's results.
- A coordinating server collects those partial updates and uses them to compare candidate network designs, selecting whichever architecture performs best across all the distributed sites.
The claim is careful to say the selection is "based, at least in part, on the type of training data" at each site, which means the architecture search is data-aware: a design that works well for radiology scans at one hospital can be weighted differently than one tuned for electronic health records at another.
Apparatuses, systems, and techniques to select a neural network architecture from a plurality of neural networks in a federated learning (FL) setting. In at least one embodiment, a neural network is trained by combining training results from different FL computing systems …
Translation: This technology picks the best AI structure by merging results from multiple computers that learn without sharing raw data.
What this means for AI trained on medical or sensitive data
Healthcare AI is the most obvious application. Training diagnostic models today usually requires either moving patient data to a central server (legally complicated and politically sensitive) or accepting that the model only sees one institution's patient population (statistically limiting). A system that can search for the best model design while respecting each institution's data boundaries could make AI diagnostics both more accurate and more broadly deployable.
The broader implication is that federated architecture search could apply anywhere data is fragmented and protected, from financial fraud detection across competing banks to predictive maintenance across manufacturing rivals. Nvidia already sells the hardware and software stacks (CUDA, Clara, Flare) that underpin these workflows, so a patent in this space fits a clear product strategy. For anyone tracking how AI infrastructure is maturing under real-world privacy constraints, this patent sits alongside a steady stream of new Big Tech patents pushing federated and privacy-preserving AI methods into specialized, regulated industries.
The engineering tradeoff here is real and worth naming: architecture search is already computationally expensive when you have all your data in one place. Distributing that search across sites with heterogeneous data types and variable network conditions multiplies the coordination overhead significantly, and the patent doesn't address how the system handles a slow or unreliable participant without biasing the architecture selection. That cost may be acceptable in high-stakes domains like medical imaging, where data simply cannot leave the building, but it makes this approach a poor fit for latency-sensitive or resource-constrained deployments. The design reads as a deliberate bet on regulated-industry customers over general-purpose ones, which is a coherent but narrow market choice.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
58 drawing sheets from US 2026/0236742 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →