Neural Network Feature Data Quantization for Mobile Facial Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks (CNNs) used in mobile devices for facial recognition face challenges in achieving robust recognition performance with limited resources, as they require significant processing time and resources, leading to degradation in recognition performance across various environments.
Innovation Solution
The method involves determining a quantization parameter based on the distribution of feature data from a neural network layer, quantizing the feature data, and inputting it to another layer, thereby reducing bit width without degrading recognition performance, improving computation and resource efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If convolutional neural networks are used for facial recognition in mobile devices, then recognition capability is improved, but processing time and resource consumption increase
Solution Approach 1:
The patent applies quantization by changing the precision parameter of neural network weights and activations from high precision (e.g., 32-bit floating point) to low precision (e.g., 8-bit integers or lower). This parameter change reduces computational complexity and memory bandwidth requirements, thereby decreasing processing time while maintaining recognition performance through careful quantization scheme design
Solution Approach 2:
The patent employs dynamic quantization techniques where quantization parameters are adjusted based on the distribution and characteristics of intermediate feature maps during network execution. This dynamic adaptation allows the system to optimize the trade-off between precision and speed for different layers and operations, reducing overall processing time while preserving recognition accuracy
2Reliability
If convolutional neural networks are used for facial recognition in mobile devices, then recognition capability is improved, but resource consumption increases
Solution Approach 1:
By changing the data precision parameter from high to low, the patent reduces memory bandwidth consumption and computational resource usage. Low-precision arithmetic operations require fewer transistors and less energy, directly reducing resource consumption while maintaining sufficient recognition performance through optimized quantization
Solution Approach 2:
The patent extracts and removes unnecessary precision from neural network computations by applying quantization. It identifies that full floating-point precision is not required for all operations and removes excess precision requirements, thereby reducing resource consumption while preserving the essential recognition capability
3Productivity
If quantization is applied to reduce bit width, then computation efficiency is improved, but recognition performance may degrade
Solution Approach 1:
The patent applies different quantization schemes to different parts of the neural network based on their specific characteristics. Critical layers that heavily impact recognition performance use higher precision or more careful quantization, while less critical layers use aggressive quantization. This local differentiation maintains overall recognition performance while achieving computation efficiency gains
Solution Approach 2:
The patent employs feedback mechanisms where quantization parameters are optimized based on measured recognition performance. Through iterative training and evaluation, the system adjusts quantization bit-widths and schemes to find the optimal balance between computation efficiency and recognition accuracy, preventing performance degradation
Data Source
AI summary
A processor-implemented neural network data processing method includes: determining a total number of either one of a first feature value and values less than or equal to the first feature value, in feature data output from a layer of a neural network; determining a quantization parameter based on the determined number; quantizing the feature data based on the determined quantization parameter; and inputting the quantized feature data to a another layer of the neural network connected to the layer.