Distributed Learning Knowledge Acquisition for Edge AI Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Discriminative AI systems struggle when training datasets cannot describe the complete scenario, while generative AI systems face issues like hallucinations and high parameter requirements.
Innovation Solution
A hybrid distributed inference system combining discriminative and generative AI models, with localized and centralized systems, where discriminative models are used at edge sites, and generative models are employed if necessary, leveraging retrieval augmented generation (RAG) and symbolic reasoning to enhance classification accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If generative AI systems are used to improve generalization ability, then the ability to handle data not in training set is improved, but the number of parameters required increases significantly
Solution Approach 1:
The patent segments the AI system into multiple smaller discriminative models, each trained on different subsets of the training data. These segmented models work together through ensemble voting to achieve generalization comparable to large generative models, but with significantly fewer parameters required at each node.
Solution Approach 2:
Each edge site in the distributed network implements local discriminative models tailored to specific local conditions and data characteristics. This local specialization allows the system to achieve strong generalization performance for local scenarios while maintaining low parameter counts at each location.
2Device complexity
If discriminative AI systems are used to reduce parameter requirements, then the number of parameters is reduced, but the ability to handle scenarios beyond training data decreases
Solution Approach 1:
The patent combines multiple discriminative models through an ensemble approach where their predictions are aggregated via voting mechanisms. This merging of multiple simple models achieves the generalization capability of complex generative models while maintaining the parameter efficiency of individual discriminative models.
Solution Approach 2:
The system implements feedback mechanisms where model predictions and performance metrics are continuously monitored and used to refine the ensemble voting strategy and individual model training, improving generalization performance over time without increasing parameter complexity.
3Measurement precision
If centralized systems are used to improve model accuracy, then classification accuracy is improved, but data transmission requirements and computational resource usage increase
Solution Approach 1:
The system segments the centralized processing function into distributed edge models that operate independently at local sites. This segmentation eliminates the need to transmit raw data to a central server, reducing data transmission requirements while maintaining classification accuracy through local inference.
Solution Approach 2:
Each edge site's local discriminative models perform self-service classification independently without requiring centralized processing. The models use local data and knowledge to make classifications autonomously, eliminating data transmission needs while achieving accurate results through the ensemble approach.
4Adaptability or versatility
If large generative models are deployed to handle complex scenarios, then the ability to generalize is improved, but computational resource requirements increase
Solution Approach 1:
The computational workload is segmented across multiple edge devices, each running lightweight discriminative models. The ensemble of segmented models collectively handles complex scenarios that would require a single large generative model, distributing computational resources efficiently across the network.
Solution Approach 2:
Instead of deploying a single large generative model, the system creates and distributes copies of smaller discriminative models to multiple edge devices. These copied models work in parallel through ensemble voting, achieving the scenario handling capability of large models with much lower computational requirements at each node.
Data Source
AI summary
The disclosure includes a distributed learning knowledge acquirer. The distributed learning knowledge acquirer allows for the training of discriminative AI models on the output data from generative AI models in a local small foundation system and a centralized large foundation system. This increases the effectiveness of the discriminative AI models which are used in an artificial intelligence hybrid distributed system to make first classifications on input data. Classifying the input data completely at the discriminative AI model level is cheaper, easier, and more time efficient making updates of the discriminative AI model critical. The distributed learning knowledge acquirer is connected to a centralized system allowing for data sharing between distributed learning knowledge acquirers in different edge sites up to and including federated learning.