Neural Network Precision Adaptation for Mobile Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network implementations require numerous operations for complex input data, leading to high power consumption and long processing times. Attempts to reduce processing requirements result in degradations in accuracy and precision.
Innovation Solution
The proposed solution involves a processor-implemented neural network method that adjusts precision based on the number of classes in the classification task. By setting higher precision for the output layer, particularly the last fully connected layer, the method minimizes accuracy loss while reducing computational requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If standard precision is used for all layers in the neural network, then processing speed and power consumption are improved, but accuracy and precision are degraded
Solution Approach 1:
The patent applies different precision levels to different layers of the neural network based on their specific requirements. The output layer uses higher precision (e.g., 32-bit floating point) to maintain classification accuracy, while intermediate layers use lower precision (e.g., 8-bit integers) to reduce computational load. This local differentiation resolves the contradiction by optimizing each layer's precision according to its functional needs rather than applying a uniform precision level throughout the network.
Solution Approach 2:
The patent dynamically adjusts the precision parameter across different layers of the neural network. By changing the precision parameter from high to low as data flows through intermediate layers, and maintaining high precision at the output layer, the system achieves both reduced computational requirements and maintained accuracy. This parameter variation strategy directly addresses the contradiction between processing efficiency and accuracy.
2Use of energy by stationary object
If processing requirements are reduced to enable deployment on mobile devices, then power consumption and processing time are improved, but accuracy and precision are degraded
Solution Approach 1:
The patent implements local quality by assigning different precision levels to different network layers based on their computational roles. Intermediate layers that perform feature extraction use lower precision to reduce power consumption, while the output layer that performs final classification uses higher precision to maintain accuracy. This resolves the contradiction by allowing power-efficient processing in most layers while preserving accuracy where it matters most.
Solution Approach 2:
The patent segments the neural network into distinct precision zones: high-precision segments for the output layer and low-precision segments for intermediate layers. This segmentation allows the system to optimize power consumption in the larger portion of the network (intermediate layers) while maintaining accuracy in the critical output segment, thereby resolving the contradiction between power efficiency and accuracy for mobile deployment.
3Productivity
If uniform low precision is applied to reduce computational load, then processing speed and power consumption are improved, but accuracy is significantly degraded
Solution Approach 1:
The patent applies local quality by differentiating precision requirements across network layers. Intermediate layers use low precision (8-bit integers) to maximize computational efficiency, while the output layer uses high precision (32-bit floating point) to ensure classification accuracy. This localized precision strategy resolves the contradiction by applying low precision only where it doesn't compromise accuracy, while maintaining high precision where it is critical for correct classification.
Solution Approach 2:
The patent changes the precision parameter across different layers of the network. By transitioning from low precision in intermediate layers to high precision in the output layer, the system achieves both computational efficiency and classification accuracy. This parameter change strategy resolves the contradiction by optimizing the precision parameter according to the functional requirements of each network layer.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A neural network method and apparatus is provided. A processor-implemented neural network method includes determining, based on a determined number of classes of input data, a precision for a neural network layer outputting an operation result, and processing parameters of the layer according to the determined precision.