Model Coefficient Decompression Using Cumulative Distribution Functions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for decompressing model coefficients in neural networks require significant memory for look-up tables, especially when the range of coefficients is large, and involve multiple tables for different subsets, leading to inefficient memory usage and increased look-ups.
Innovation Solution
The use of decompression functions based on cumulative count distributions, where function parameters are stored and applied to compressed model coefficients, reducing memory consumption and look-ups by allowing a single set of parameters to decompress multiple coefficients.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If look-up tables are used for decompressing model coefficients, then decompression accuracy is maintained, but memory consumption increases significantly
Solution Approach 1:
The patent extracts only the essential statistical parameters (mean, standard deviation, minimum, maximum) from the full coefficient distribution, storing only these extracted features rather than the complete distribution data. This allows accurate reconstruction of coefficient values through mathematical formulas while using minimal memory.
Solution Approach 2:
The patent replaces the mechanical look-up table system with a mathematical function-based system. Instead of storing pre-computed values in tables, the system uses parametric distributions (Gaussian, uniform, etc.) with stored parameters to generate coefficient values through mathematical equations, significantly reducing memory requirements.
2Measurement precision
If multiple look-up tables are used for different coefficient subsets, then decompression precision is improved, but device complexity increases
Solution Approach 1:
The patent creates a universal decompression framework that handles multiple coefficient subsets using a single set of mathematical functions and parameter storage. The system can adapt to different distribution types (Gaussian, uniform, etc.) and different coefficient ranges through parameter variation rather than requiring separate look-up tables for each subset.
Solution Approach 2:
The patent changes the approach from storing fixed look-up table data to storing adjustable parameters (mean, standard deviation, min, max) that define parametric distributions. This allows the same mathematical framework to handle different coefficient subsets by simply changing the stored parameters rather than using multiple tables.
3Quantity of substance
If compressed coefficients are stored to reduce memory usage, then memory efficiency improves, but decompression time increases due to multiple look-ups
Solution Approach 1:
The patent performs preliminary statistical analysis during the compression phase to calculate and store the essential parameters (mean, standard deviation, minimum, maximum) of coefficient distributions. This preliminary action enables fast decompression later, as the parameters are already computed and stored, eliminating the need for multiple look-ups during the decompression phase.
Data Source
AI summary
A predictive model utilizes a set of coefficients for processing received input data. To reduce memory usage storing the coefficients, a compression circuit compresses the set of coefficients prior to storage by generating a cumulative count distribution of the coefficient values, and identifying a distribution function approximating the cumulative count distribution. Function parameters for the determined function are stored in a memory and used by a decompression circuit to apply the function the compressed coefficients to determine the decompressed component values. Storing the function parameters may consume less memory in comparison to storing a look-up table for decompression, and may reduce an amount of memory look-ups required during decompression.