Sony Patents a Way to Auto-Design AI Networks Without Needing Real Training Labels
Training an AI usually requires mountains of carefully labeled data. Sony is patenting a way to skip that bottleneck when designing the AI's architecture in the first place.
What Sony's pseudo-label AI design system actually does
Every time an AI model gets built, engineers face two expensive problems: choosing the right network shape, and collecting enough labeled examples to train it. Labeling data, where humans tag images, audio clips, or text with correct answers, can cost months and millions of dollars.
Sony's patent describes a system that sidesteps part of that cost. Instead of waiting for real labels, it attaches fake, generated labels (called pseudo-labels) to raw data, feeds that stand-in dataset into an algorithm that automatically tests different AI designs, and checks how the algorithm is coping. If the error score stays high after that test, the system flags a particular AI structure as worth using.
In plain terms: Sony's approach lets a machine test-drive different AI blueprints using synthetic stand-in answers, rather than real human-verified ones. That could make the early design phase of building AI faster and cheaper.
a generation unit that assigns a pseudo label, which is a pseudo label, to data, so as to generate a pseudo data set including the data assigned the pseudo label; …
Translation: The system automatically labels unlabelled data to create a custom training dataset.
How the loss-threshold check selects a network structure
The patent describes three linked components working together inside an information processing apparatus.
- Generation unit: Takes unlabeled data and stamps each item with a pseudo-label, a computer-generated guess at what the correct answer should be. This synthetic dataset stands in for expensive human-annotated training data.
- Determination unit: Feeds that pseudo-dataset into a neural architecture search (NAS) algorithm. NAS is a technique where software automatically hunts through thousands of possible AI network shapes to find one that works best, rather than having engineers design the network by hand. The unit watches the algorithm's loss value (a number measuring how wrong the model's predictions are). If that error number meets or exceeds a set threshold, the system treats it as a meaningful signal.
- Output control unit: When the loss threshold is triggered, it outputs structure information, a description of the neural network layout the NAS algorithm was evaluating at that point.
The logic is counterintuitive: a high loss with fake labels signals that the network structure is sensitive and expressive enough to be useful, making the error score a proxy for architectural quality rather than a sign of failure.
… an output control unit that outputs structure information regarding a structure of a neural network, in a case where the determination unit determines that the loss value is equal to or greater than the predetermined threshold.
Translation: It finally exports the newly designed AI architecture once the error metrics hit the target.
What this means for building AI without expensive labeled data
For companies building AI at scale, the architecture-search phase is one of the most resource-intensive steps. If Sony's approach works, engineers could run those searches on raw, unlabeled data they already have, rather than waiting for annotated datasets to arrive. That trims both time and cost from a step that currently requires significant human labor.
For everyday users, the downstream effect would be AI products that reach the market faster and potentially at lower cost. Sony's steady investment in AI training methods also fits a pattern of hardware-adjacent companies building proprietary AI infrastructure, which often shows up later in cameras, audio gear, and gaming systems.
This is the sixth Sony filing we've tracked since July in our on-device AI privacy watchlist, building on ideas like training on approximate private data and split training over radio links.
Claim 1 covers any device that attaches a guessed label to data, checks whether a quality score during an automated AI-design process meets a threshold, and then delivers a blueprint for a neural network. The claim places no limits on the type of data, the labeling method, or the automated design process involved, which makes the fence around it very wide.
That width has real consequences. Almost any software pipeline that uses estimated or invented labels to steer an automated AI-building process could fall inside this claim, across industries from medical imaging to consumer apps.
The central question an examiner will press is whether combining two already-common practices, guessing at labels and automating network design, into a single workflow earns patent protection on its own. The claim language does not answer that question, and everything hinges on how strictly that standard gets applied.
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
6 drawing sheets from US 2026/0268144 A1 · click any drawing to enlarge
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →