Shared Weight Parameters for Multi-Model Edge AI Memory Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods fail to reduce memory capacity when implementing multiple learned models on devices due to the need for storing multiple sets of weight parameters, which is particularly challenging for edge devices with limited memory.
Innovation Solution
Performing a first and second learning process using different data sets to generate models with shared weight parameters by updating them to have almost the same values through an arbitration mechanism.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple learned models are stored separately on an edge device, then different learning tasks can be performed, but the memory capacity required increases significantly
Solution Approach 1:
The patent merges multiple learned models into a single shared model structure where different learning tasks share common weight parameters. By making weight parameters common across multiple models and only maintaining task-specific parameters separately, the system reduces memory requirements while preserving the ability to perform different learning tasks.
Solution Approach 2:
The patent creates a universal base model that can serve multiple learning tasks simultaneously. The shared weight parameters form a universal foundation that can be adapted to different tasks through task-specific parameters, allowing one model structure to fulfill multiple functions rather than requiring separate dedicated models for each task.
2Quantity of substance
If weight parameters are deleted through pruning to reduce model size, then memory capacity is reduced, but recognition accuracy deteriorates
Solution Approach 1:
The patent extracts and separates task-specific parameters from the shared weight parameters. By taking out only the necessary task-specific components and leaving the shared parameters intact, the system achieves model compression without removing critical weight parameters that would harm recognition accuracy.
Solution Approach 2:
The patent changes the parameter organization from separate full models to a hybrid structure with shared and task-specific parameters. This parameter reorganization allows the system to maintain sufficient parameters for accurate recognition while reducing overall memory requirements through efficient parameter sharing.
3Quantity of substance
If transfer learning is used to reduce memory capacity, then memory efficiency improves, but the method cannot be applied when learned models are generated through relearning of entire layers
Solution Approach 1:
The patent introduces a dynamic parameter sharing mechanism that can adaptively determine which parameters to share and which to keep task-specific. This dynamic approach allows the system to flexibly adjust the degree of parameter sharing based on the specific learning scenario, making it applicable to various situations including relearning of entire layers where traditional transfer learning would fail.
Data Source
AI summary
An information processing device 500 of the present invention includes a first learning means 521 for performing a first learning process by using a first learning data set to generate a first learning model including a first weight parameter, a second learning means 522 for performing a second learning process by using a second learning data set to generate a second learning model including a second weight parameter, and an arbitration means 523 for, when performing the first learning process and the second learning process, updating the first weight parameter and the second weight parameter such that the values of the first weight parameter and the second weight parameter become almost the same.