Neural Network Parameter Coding With Outlier-Aware Bit Truncation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
As artificial neural networks generate a large number of parameters during learning, training, or inference, the memory devices required to store these parameters consume significant power, necessitating a reduction in power consumption without compromising accuracy.
Innovation Solution
An electronic device with a coding module that determines whether a parameter is an outlier based on its value and compresses it by truncating specific bits, with the decoding module restoring the compressed parameters, thereby reducing memory storage needs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If parameters are stored with full precision in memory devices, then accuracy of artificial neural network is maintained, but power consumption of memory device increases
Solution Approach 1:
The patent applies parameter changes by modifying the precision representation of neural network parameters through bit truncation. The coding module truncates least significant bits of parameter values to reduce memory storage requirements and power consumption while maintaining acceptable accuracy for neural network operations
Solution Approach 2:
The patent implements local quality by differentiating the treatment of parameters based on their statistical properties. Outlier parameters (those with absolute values greater than threshold) are truncated by one bit, while non-outlier parameters are truncated by two bits, optimizing the balance between accuracy and power consumption for different parameter types
2Use of energy by stationary object
If parameters are compressed by truncating bits, then power consumption of memory device is reduced, but accuracy of artificial neural network deteriorates
Solution Approach 1:
The patent uses parameter changes by dynamically adjusting the truncation level based on parameter characteristics. The coding module changes the number of bits truncated depending on whether each parameter is an outlier or not, thereby controlling the degradation of accuracy while achieving power consumption reduction
Solution Approach 2:
The patent applies partial action by selectively truncating only certain bits of parameters rather than all bits. This partial truncation approach reduces power consumption to an acceptable level while maintaining sufficient accuracy for neural network functionality
3Device complexity
If uniform bit truncation is applied to all parameters, then implementation complexity is reduced, but accuracy loss increases for outlier parameters
Solution Approach 1:
The patent implements local quality by applying different truncation strategies to different parameter types. The coding module identifies outlier parameters and applies a less aggressive truncation (one bit) compared to non-outlier parameters (two bits), thereby preserving accuracy for critical parameters while maintaining relatively simple implementation
Data Source
AI summary
An electronic device includes a coding module that determines whether a parameter of an artificial neural network is an outlier, depending on a value of the parameter and compresses the parameter by truncating a first bit of the parameter when the parameter is a non-outlier and truncating a second bit of the parameter when the parameter is the outlier, and a decoding module that decodes a compressed parameter.