Adaptive Neural Network Weight Reduction for Electronic Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network weight reduction methods are not scalable and inefficiently applied to various electronic device scenarios, leading to uniform operation that does not consider resource variations.
Innovation Solution
An electronic device that performs neural network weight reduction operations based on resource information, allowing for adaptive selection of weight information classes to optimize memory and power usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If neural network weight reduction is uniformly applied according to device resources, then memory bandwidth and power consumption are reduced, but scalability and efficiency for various device situations are lost
Solution Approach 1:
The patent implements dynamic weight reduction by determining reduction parameters based on real-time resource information of the electronic device. The weight reduction operation adapts to varying device conditions (CPU resources, memory capacity, power state) rather than using a fixed uniform approach, enabling the system to optimize performance for different device scenarios while maintaining scalability.
Solution Approach 2:
The system changes weight reduction parameters dynamically based on resource information. Different weight reduction operations are selected according to available CPU resources, memory capacity, and power state, allowing the same neural network to be optimized for different device configurations without requiring separate models for each scenario.
2Device complexity
If neural network weight reduction is uniformly applied, then implementation simplicity is maintained, but operational efficiency for various device situations deteriorates
Solution Approach 1:
The weight reduction operation is segmented into multiple types (first weight reduction operation and second weight reduction operation) with different reduction parameters. The system selects appropriate segments based on device resource information, maintaining implementation simplicity through modular design while achieving operational efficiency by applying the most suitable segment to each device situation.
Solution Approach 2:
The system dynamically selects which weight reduction operation to apply based on real-time resource assessment. This dynamic selection mechanism maintains relatively simple implementation while significantly improving operational efficiency by adapting to varying device conditions such as CPU availability, memory capacity, and power state.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An electronic device is provided. The electronic device includes: a memory configured to store first information and second information; and a processor configured to obtain a first weight matrix by loading the first information or obtain a second weight matrix by loading the first information and the second information, based on resource information of the electronic device, wherein the first information includes weights related to the first weight matrix and a first index corresponding to the weights, and the second information includes an additional weight for obtaining the second weight matrix and a second index corresponding to the additional weight.