Neural Network Quantization Bit Alignment for Recognition Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks quantized using existing methods often experience a degradation in recognition accuracy during the learning process, particularly in image processing tasks, due to the complexity of computation processing and the resulting increased number of computations and memory requirements.
Innovation Solution
An information processing apparatus and method that includes a processor and memory configured to perform quantization on neural network variables by obtaining the distribution of appearance frequencies and aligning the most significant bit position of the quantization position to a variable element smaller than the maximum value, thereby reducing the impact of quantization on recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If quantization is applied to reduce computation complexity, then execution time is reduced, but recognition accuracy degrades
Solution Approach 1:
The patent changes the parameter of quantization position (bit alignment) from conventional fixed alignment to dynamic alignment based on appearance frequency distribution. By adjusting which bit position aligns with the most significant bit according to the actual data distribution, the patent maintains higher precision in representing frequently occurring values while still using fixed-point quantization for efficiency.
Solution Approach 2:
The patent applies local quality by treating different bit positions differently based on the frequency distribution of variable elements. Instead of uniform quantization across all bit positions, the patent aligns the quantization grid with the actual distribution characteristics, giving higher precision to the most significant bits that carry the most information while accepting coarser quantization for less significant bits.
2Reliability
If computation processing is complicated to achieve high performance, then recognition accuracy is improved, but the number of computations and memory capacity increase
Solution Approach 1:
The patent changes the parameter of quantization representation from conventional formats to a format optimized for the specific data distribution. By modifying how values are quantized and represented based on their appearance frequency distribution, the patent achieves efficient computation with reduced complexity while maintaining the accuracy needed for high-performance recognition tasks.
3Quantity of substance
If quantization position is aligned to maximum value variable element, then representation range is maximized, but precision for frequently occurring smaller values decreases
Solution Approach 1:
The patent changes the alignment parameter of quantization from fixed alignment to dynamic alignment based on appearance frequency distribution. This allows the quantization grid to be optimally positioned to capture the precision needs of frequently occurring values while still accommodating the full range of possible values through appropriate scaling and offset adjustments.
Solution Approach 2:
The patent applies local quality by concentrating precision where it is most needed - on the most significant bits corresponding to frequently occurring values. The quantization scheme provides higher resolution for common values while maintaining adequate representation for rare extreme values, optimizing the trade-off between precision and range.
Data Source
AI summary
An information processing apparatus includes: a memory configured to store program instructions to perform quantization on quantization target data; and a processor configured to execute the program instructions stored in the memory, the program instructions including: obtaining a distribution of appearance frequencies of a plurality of variable elements included in the quantization target data; and aligning a most significant bit position of a quantization position to a variable element smaller than a variable element of a maximum value among the plurality of variable elements based on the distribution of the appearance frequencies of the plurality of variable elements.