IoT Model Tiering for Prediction Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
IoT devices face challenges in maintaining prediction accuracy due to limited computational power and changing data types, leading to decreased reliability over time, especially when encountering novel conditions not covered in their initial training data.
Innovation Solution
Implementing a split prediction system where a local model on a hub device generates an initial prediction, which is then corrected by a more accurate prediction from a provider network, and periodically updating local models based on new data to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If a simple data processing model is implemented on an IoT device to enable local prediction, then the device can perform predictions independently, but the reliability and accuracy of the prediction deteriorates compared to larger models on more powerful computing devices
Solution Approach 1:
The prediction system is segmented into two parts: a simple local model on the IoT device for quick independent predictions, and a larger remote model on the provider network for high-accuracy predictions. This segmentation allows each component to serve its specific function optimally without requiring the IoT device to host the entire large model.
Solution Approach 2:
The patent introduces an intermediary mechanism where the simple local model's predictions are compared with predictions from the larger remote model. The remote model acts as a mediator that corrects or validates the local predictions, thereby improving overall reliability while maintaining the benefit of local inference capability.
2Reliability
If a large model with hundreds of millions of parameters is used to improve prediction accuracy, then the reliability improves, but the computational power and resources required increase significantly
Solution Approach 1:
The model is segmented into a small local model for deployment on resource-constrained IoT devices and a large remote model hosted on powerful server infrastructure. This segmentation allows the computational burden of the large model to be offloaded to the provider network while retaining the accuracy benefits.
Solution Approach 2:
Instead of deploying the full large model to every IoT device, a simplified copy or distilled version is deployed locally. The local copy performs quick predictions, and discrepancies are resolved by querying the full remote model, thus reducing local computational requirements while maintaining accuracy.
3Reliability
If the local model is trained initially with adequate training data to achieve good accuracy, then the prediction reliability is improved, but the model loses accuracy over time when encountering novel conditions not covered in training data
Solution Approach 1:
The system implements a feedback mechanism where predictions from the local model are compared with predictions from the remote model. When discrepancies are detected or when the remote model encounters novel conditions, the results are used to update and retrain the local model, enabling it to adapt to new conditions while maintaining its initial accuracy on known data.
Solution Approach 2:
The local model is preliminarily trained with adequate training data to achieve good initial accuracy. This preliminary training ensures the model performs well on known conditions, while the feedback loop with the remote model prepares the system to handle novel conditions that arise later.
4Adaptability or versatility
If the model parameters are increased to handle more data types and conditions, then the adaptability improves, but the device complexity and resource requirements increase
Solution Approach 1:
The remote model on the provider network serves as a universal handler for diverse data types and conditions. The local IoT device only needs to implement a simple model, while the remote universal model handles the complexity of multiple data types, thus achieving versatility without increasing local device complexity.
Data Source
AI summary
Edge devices of a network collect data. An edge device may determine whether to process the data using a local data processing model or to send the data to a tier device. The tier device may receive the data from the edge device and determine whether to process the data using a higher tier data processing model of the tier device. If the tier device determines to process the data, then the tier device processes the data using the higher tier data processing model, generates a result based on the processing, and sends the result to an endpoint (e.g., back to the edge device, to another tier device, or to a control device). If the tier device determines not to process the data, then the tier device may send the data on to another tier device for processing by another higher tier model.