Channel-Wise Data Quantization for Neural Network Precision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
As neural network complexity increases, the data volume and dimension also increase, posing challenges to data processing efficiency, storage capacity, and access efficiency. Existing quantization schemes for whole neural networks often lead to lower precision and affect data operation results due to differences in operational data.
Innovation Solution
A data quantization processing method and apparatus that determines data to be quantized for each channel of an input image in a deep convolution layer, calculates statistical results for each channel, and uses these results to determine channel-specific quantization parameters for precise quantization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a same quantization scheme is adopted for a whole neural network, then device complexity is reduced, but measurement precision deteriorates due to large differences between different operational data
Solution Approach 1:
The patent divides the neural network into multiple channels and applies different quantization parameters to each channel. Specifically, for each channel in the deep convolution layer, separate statistical results (mean and standard deviation) are calculated, and distinct quantization parameters are determined based on these channel-specific statistics. This segmentation approach resolves the contradiction by maintaining low overall system complexity while achieving high precision through channel-level differentiation.
Solution Approach 2:
The patent implements local quality by adapting quantization parameters to the specific characteristics of each channel. Instead of using a uniform quantization scheme across the entire network, the method calculates channel-specific statistical properties and derives customized quantization parameters for each channel. This allows each local region (channel) to have optimized quantization properties matched to its data distribution, thereby improving measurement precision without significantly increasing device complexity.
2Adaptability or versatility
If data volume and dimension increase to accommodate neural network complexity, then adaptability improves, but productivity deteriorates due to challenges in data processing efficiency and storage access
Solution Approach 1:
The patent addresses the productivity challenge by changing the parameter representation from high-precision floating-point to low-precision fixed-point quantized values. By determining quantization parameters based on channel-specific statistical results and applying these parameters to convert operational data, the system maintains the adaptability needed for complex neural networks while significantly improving data processing efficiency and reducing storage access challenges through compact quantized data representation.
Data Source
AI summary
The present disclosure relates to a data quantization processing method and apparatus, an electronic device, and a storage medium. The apparatus includes a control unit having an instruction caching unit, an instruction processing unit, and a storage queue unit. The instruction caching unit is configured to store a calculation instruction associated with an artificial neural network operation, the instruction processing unit is configured to parse the calculation instruction to obtain a plurality of operation instructions, and the storage queue unit is configured to store an instruction queue. The instruction queue includes a plurality of operation instructions or calculation instructions to be executed in an order of the queue. The above-mentioned method improves the operation precision of related products during a neural network model operation.