Edge AI Model Conversion Behavior Alignment via Cross-Model Error Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing prediction models converted for lighter execution on edge devices often exhibit different behaviors despite maintaining prediction performance, leading to inconsistent results, particularly in safety-critical applications like object recognition, where differences can cause safety risks.
Innovation Solution
An information processing method that calculates prediction results from both original and converted models using different data types, calculates errors, and trains the converted model using machine learning to align its behavior with the original model, thereby maintaining or improving recognition performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If prediction model conversion is performed to make processing lighter for edge devices, then processing efficiency and resource consumption are improved, but prediction behavior consistency deteriorates
Solution Approach 1:
The patent applies feedback by calculating prediction results from both the original and converted models, computing errors between them, and using these errors to retrain the converted model. This feedback loop ensures the converted model's behavior aligns with the original model while maintaining processing efficiency for edge devices.
Solution Approach 2:
The patent changes parameters by retraining the converted model using machine learning based on calculated errors. This involves adjusting model parameters through additional training iterations to minimize behavioral differences while preserving the lightweight structure needed for efficient edge device execution.
2Weight of stationary object
If prediction model conversion is performed to reduce model size, then device resource consumption is reduced, but individual prediction result consistency deteriorates
Solution Approach 1:
The system calculates prediction results from both original and converted models, computes errors between corresponding predictions, and uses this feedback to retrain the converted model. This ensures individual prediction results remain consistent despite the reduced model size.
Solution Approach 2:
The patent performs preliminary actions by pre-calculating errors between original and converted model predictions before final training. This preliminary error calculation guides the subsequent training process to focus on correcting specific prediction discrepancies.
3Productivity
If statistical prediction results are maintained between models, then overall performance is preserved, but individual behavior alignment deteriorates
Solution Approach 1:
The patent changes parameters by retraining the converted model using machine learning based on calculated errors. This involves adjusting model parameters through additional training iterations to minimize behavioral differences while preserving the lightweight structure needed for efficient edge device execution.
Data Source
AI summary
An information processing includes: obtaining first data belonging to a first type and second data belonging to a second type different from the first type; calculating a first prediction result by inputting the first data into a first prediction model; calculating a second prediction result by inputting the first data into the second prediction model; calculating a third prediction result by inputting the second data into the second prediction model; calculating a first error between the first prediction result and the second prediction result; calculating a second error between the second prediction result and the third prediction result; and training the second prediction model by machine learning, based on the first error and the second error.