Neural Network Quantization With Adaptive Scale Parameters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing hardware deep neural network (DNN) accelerators face high computing and memory requirements for tasks like image enhancement and super-resolution, and existing quantization methods often result in performance degradation.
Innovation Solution
The proposed solution involves a shifter and decoder configuration that uses a scale parameter to minimize quantization error by selecting data subsets with high similarity to those generated by the Lloyd-Max quantization technique, allowing for iterative optimization of quantization scales for each channel or layer of a neural network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If quantization is applied to reduce hardware cost and memory space, then hardware complexity is reduced, but quantization error increases causing performance degradation
Solution Approach 1:
The patent applies parameter changes by introducing a scale parameter that dynamically adjusts the quantization process. The scale parameter modifies the quantization intervals based on the statistical properties of the input data, allowing the system to maintain higher precision where needed while still achieving hardware reduction. This is implemented through iterative optimization that adjusts the scale parameter to minimize quantization error for each layer or channel independently.
Solution Approach 2:
The patent implements local quality by applying different quantization strategies to different parts of the neural network. Specifically, different scale parameters are applied to different layers or channels based on their individual characteristics. This allows each layer to have optimized quantization parameters tailored to its specific data distribution, rather than using a uniform quantization approach across the entire network.
2Device complexity
If uniform quantization scale is used for all channels and layers, then device complexity is reduced, but quantization accuracy deteriorates
Solution Approach 1:
The patent applies segmentation by dividing the neural network into distinct groups based on layer type or channel characteristics. Each group is then assigned its own scale parameter, allowing independent optimization. This segmentation approach balances the complexity of managing multiple parameters with the benefit of improved quantization accuracy for each specific group.
Solution Approach 2:
The patent implements dynamics by making the quantization scale adaptive rather than static. The scale parameter is determined through iterative optimization processes that adjust it based on the actual data distribution and quantization error feedback. This dynamic adjustment allows the system to optimize quantization accuracy without requiring complex manual configuration.
3Measurement precision
If iterative optimization is performed to minimize quantization error, then quantization accuracy is improved, but computation time increases
Solution Approach 1:
The patent applies preliminary action by pre-computing statistical properties of the data (such as mean and standard deviation) before the actual quantization process. These pre-computed statistics are then used to initialize the scale parameter, significantly reducing the number of iterative optimization steps needed. This preliminary preparation phase enables faster convergence during the actual quantization optimization.
Data Source
AI summary
An electronic device includes: a shifter configured to perform a shift operation based on a codebook supporting a plurality of quantization levels preset for data bits of a data set; and a decoder configured to control the shifter by setting quantization scales of the data bits differently for preset groups, wherein the shifter is configured to quantize and output the data bits by control of the decoder.