Pose the question most people are too polite to ask: why would you mess up your own training data on purpose? Because a model that has only ever seen a tumor centered, upright, and well-lit will fail the moment it sees one that's off-center or dim. Transforming the data forces the model to learn the thing, not the framing.

US20200293828A1 (published September 2020) describes generating transformed versions of training inputs and using them to train the network, squarely in CPC G06N 3/08 and G06N 3/063, the neural-network learning and hardware-implementation classes. The method is general, but the inventor team's other filings cluster in medical imaging, where labeled data is scarce and augmentation pays off most.

“Apparatuses, systems, and techniques to perform training of neural networks using stacked transformed images.”— U.S. Patent Application 2020/0293828 A1 source

The disclosure is more pointed than generic augmentation, and the abstract says why: the network is trained on stacked transformed images and then “provided to be used for processing images from an unseen domain distinct from a source domain, wherein stacked transformed images are transformed according to transformation aspects related to domain variations.” The target here is not just robustness to wobble, it is domain transfer. You train on images from one domain, deliberately transform them to anticipate the ways a second domain will differ, and the resulting model works on images it was never trained on.

The first claim spells out the structure: one or more circuits help train a first set of neural networks on a first set of images, drawn from a first domain, to identify objects in a second set of images drawn from a second domain, “wherein the first set of images are transformed prior to training based on expected differences between the first domain and the second domain.” That phrase, expected differences between domains, is the whole insight. The transformations are not random jitter; they are chosen to model the specific gap between where your labeled data came from and where the model must actually perform. A dependent claim adds the hardware framing, a “transformer” component that takes a source image and produces a transformed image according to an image aspect, plus storage for both, reflecting the G06N 3/063 hardware-implementation tag.

Under the hood, the mechanism is a multiplier on your dataset. Ten thousand labeled images become effectively hundreds of thousands of training views, each teaching the same lesson from a different angle. The model can't memorize a specific pixel arrangement, so it has to generalize. But the “stacked” and “domain” language points at something sharper than volume: by stacking transformed versions and tying the transformations to expected domain shift, the method tries to pre-bake the destination domain's characteristics into the source-domain training, so the model arrives already adapted.

This connects to a recurring neuraldocket theme: the least glamorous step often decides whether a model works. We've covered NVIDIA's own data-augmentation IP in a separate piece; this 2020 publication is an earlier point on the same curve, showing the technique was standard practice and worth protecting before the foundation-model era, and that the framing had already advanced from “make the model robust” to “make the model port across domains it has never seen.”

Reading the claim language closely, the method is engineered around a known gap rather than generic robustness. The training images come from a “first domain”; the images the model must eventually classify come from a “second domain”; and the transformations are chosen “based on expected differences between the first domain and the second domain.” That is domain adaptation done preemptively, at the data layer, before a single training step runs. If you know that your labeled scanner produces brighter, higher-contrast images than the deployment scanner, you transform the training set to anticipate that shift, so the model never has to confront the gap at inference time as a surprise. The “stacked” framing in the abstract suggests assembling multiple such transformed versions together as the training input, multiplying the variety the network sees.

The dependent claim's hardware framing is a tell about where this was meant to run. It recites first storage for the source images, a dedicated “transformer” component that produces a transformed image according to an image aspect, and second storage for the result, all consistent with the G06N 3/063 hardware-implementation class. In other words, the transformation is not an afterthought in a data-loading script; the filing contemplates it as a pipeline stage with its own silicon and memory, fast enough to feed a GPU training loop without becoming the bottleneck. That is the unglamorous engineering that lets augmentation scale to the millions of views a large model demands.

Placing this on the neuraldocket timeline sharpens what it does and does not claim. By September 2020 the broad idea of data augmentation was decades old and unpatentably generic; what the filing stakes out is the narrower, more defensible territory of transformation-as-domain-adaptation, transformations selected to bridge a specific, known gap between a labeled source domain and an unlabeled deployment domain, implemented as a hardware pipeline stage. That is a meaningfully different claim from “flip and rotate your images.” It treats augmentation as a controlled instrument for moving a model from where its data came from to where it has to work, and it situates that instrument in silicon fast enough to keep a training loop fed. The honesty of the limitation, that the whole thing depends on choosing transformations that actually mirror the domain shift and preserve the label, is also what keeps it from overclaiming.

The honest gloss on the limits: augmentation helps when the transformations match real-world variation and hurts when they don't. Flip a chest X-ray left-to-right and you might teach the model the heart is on the wrong side. The patent's own domain-difference framing concedes the catch, the transformations must genuinely reflect the expected gap between domains, or they teach the wrong lesson. The art is choosing transformations that preserve the label and model the right shift. The patent describes the machinery; the judgment stays human.