Training neural networks using consistency measures

This invention describes a method for training a main artificial intelligence (AI) system by having it process information for a task and then comparing its output to what other AI systems predict for related tasks using the same information. If the main AI system's output doesn't align with these 'consistency targets' from the other AIs, it learns from that difference. This approach is especially useful for training with data that doesn't have explicit answers, though it can also incorporate labeled data, and specifically involves multiple neural networks working together.

Why it matters: Filed before the widespread adoption of large foundation models, this method for training with consistency measures and unlabeled data is now highly relevant for efficiently fine-tuning or adapting such models to new tasks without extensive new labeling. The increasing complexity and data demands of modern AI systems make techniques that leverage consistency across tasks more valuable.

Status: Not in forceListed as no longer active. The specific reason is not in the record we hold.
How hard to build: SpecializedComplex software, multiple neural networks

3 AI reinvention directions for this patent

See all three

AI gives you a few directions you could take this. Pick one, and we check whether your version is different enough to patent, then write the filing.

Reinvent this with AI

Where this one came from

Every morning we go through millions of abandoned patents and pick out 15 worth your time. This was one of today’s.

This one

+10