Neural Network Log Quantization With Error-Corrected Parameters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks require efficient computation methods to process complex input data in real-time, particularly in low-power and high-performance embedded systems like smartphones, where resources are limited.
Innovation Solution
A method involving log quantization to determine first and second quantization values for neural network parameters, grouping them with tag bits and code values, and performing convolution operations to maintain accuracy while reducing computation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If neural network parameters are quantized to reduce data size, then data storage requirements are reduced, but recognition accuracy deteriorates
Solution Approach 1:
The patent segments the quantization process into multiple stages: first quantization followed by error calculation and second quantization. This segmentation allows the system to separately handle the coarse quantization (for data size reduction) and the error correction (for accuracy maintenance), resolving the contradiction between compression ratio and recognition accuracy.
Solution Approach 2:
The patent changes the quantization parameter representation by introducing a multi-value system where parameters can be represented by first quantization values, second quantization values, or combinations thereof. This parameter change enables flexible representation that adapts to different accuracy requirements while maintaining data compression, thus resolving the trade-off between data size and recognition accuracy.
2Quantity of substance
If log quantization is applied to neural network parameters, then data compression efficiency is improved, but computational complexity increases
Solution Approach 1:
The patent performs preliminary log quantization to generate first quantization values before the main processing stage. This preliminary action prepares the data in a compressed format that facilitates subsequent error calculation and second quantization, improving overall compression efficiency while organizing the computational workflow to manage complexity systematically.
Solution Approach 2:
The patent introduces an intermediary error value that mediates between the first quantization result and the final parameter representation. This intermediary element allows the system to calculate and compensate for quantization errors without requiring complete re-quantization, thus improving compression efficiency while keeping computational complexity manageable through targeted error correction.
3Reliability
If multiple quantization values are used per parameter, then recognition accuracy is maintained, but data structure complexity increases
Solution Approach 1:
The patent implements a nested data structure where first quantization values and second quantization values are hierarchically organized. The first quantization value serves as the base layer, and the second quantization value is nested as an optional error correction layer. This nesting allows the system to maintain recognition accuracy through multi-value representation while managing data structure complexity through a systematic hierarchical organization.
Solution Approach 2:
The patent applies partial action by selectively applying second quantization values only when needed for error correction, rather than uniformly applying full multi-value representation to all parameters. This selective approach maintains recognition accuracy for parameters that require it while avoiding unnecessary data structure complexity for parameters where first quantization suffices, thus resolving the contradiction between accuracy and structural simplicity.
Data Source
AI summary
A processor-implemented method includes determining a first quantization value by performing log quantization on a parameter from one of input activation values and weight values in a layer of a neural network, comparing a threshold value with an error between a first dequantization value obtained by dequantization of the first quantization value and the parameter, determining a second quantization value by performing log quantization on the error in response to the error being greater than the threshold value as a result of the comparing; and quantizing the parameter to a value in which the first quantization value and the second quantization value are grouped.