How AI Models Learn from Both Local and Central AI Systems

This patent describes a system for training specialized AI models at local sites, like factories, by combining insights from small local AI systems and large central AI systems, especially when data patterns change.

What it covers

The patent describes a "distributed learning knowledge acquirer" which is a system designed to improve how AI models learn. This acquirer sits at a "distributed edge site," like a factory (Claim 2), and gathers data from three places: a "local small foundation system" (a smaller AI model at the edge), a "distributed endpoint management system" (which manages devices at the edge), and a "centralized large foundation system" (a powerful AI model in a central location). The acquirer checks if this incoming data shows a "conceptual drift" (a change in data patterns). If it does, this new data is used as "local training data" to update a "local discriminative model." This local model then learns from both the local small AI's classifications and the central large AI's classifications (Claim 1), making it better at quickly classifying new input data at the edge site (Claim 3).

What it doesn't cover

• —Does not cover training discriminative AI models without using data from both a local small foundation system and a centralized large foundation system.
• —Does not cover training AI models that do not perform a "drift evaluation" to identify data causing "conceptual drift" before using it for training.
• —Does not cover systems where the local small foundation system has more than 10 billion parameters (Claim 5).
• —Does not cover training discriminative AI models solely on human-labeled data, rather than on output data from other AI models.
• —Does not cover systems that lack a "distributed learning knowledge acquirer" to coordinate data reception and model training across edge and central systems.

The clever bit

The novelty lies in intelligently combining the strengths of small, local generative AI models with large, centralized generative AI models to train discriminative AI models. It specifically uses the concept of "conceptual drift" to identify when local data changes enough to warrant updating the local model, ensuring the edge AI remains relevant and accurate.

Why it matters

In today's world, many devices, like those in smart factories, need to make quick decisions without sending all their data to a central cloud. This patent helps train "discriminative AI models" (AI that classifies things) more effectively and cheaply right where they are used. By combining local, specialized knowledge with broad, central intelligence, these models can adapt quickly to new situations, like changes on a factory floor, making the entire system more efficient and responsive.

Real-world examples

• 1.Quality control AI in smart factories
• 2.Predictive maintenance systems on industrial equipment
• 3.Edge computing devices for real-time data analysis
• 4.Automated inspection systems in manufacturing
• 5.Smart city infrastructure monitoring