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 during decompression.
Innovation Solution
The use of decompression functions based on cumulative count distributions, such as bimodal, Gaussian, or Poisson distributions, where function parameters are stored and used by a decompression circuit to decompress compressed model coefficients, reducing memory requirements and look-ups by allowing a single set of function 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 can be performed, but significant memory is required especially when the range of coefficients is large
Solution Approach 1:
The patent transforms the decompression approach from using fixed look-up tables to using parametric functions (Gaussian, Poisson, or uniform distributions) characterized by a small set of parameters (mean, variance, rate, etc.). This parameterization allows the same function with different parameters to decompress coefficients across different ranges and distributions, dramatically reducing memory requirements while maintaining decompression accuracy.
Solution Approach 2:
The patent applies different distribution functions (Gaussian, Poisson, uniform) to different subsets of coefficients based on their local statistical characteristics. By identifying the appropriate distribution type for each coefficient subset and using corresponding parametric functions, the system achieves accurate local decompression without requiring large global look-up tables.
2Adaptability or versatility
If multiple look-up tables are used for different subsets of coefficients, then comprehensive coverage is achieved, but the number of look-ups during decompression increases
Solution Approach 1:
The patent creates universal decompression functions that can handle multiple coefficient distributions and ranges through parameter adjustment. A single Gaussian function, for example, can decompress coefficients across the entire range by adjusting its mean and variance parameters, eliminating the need for multiple specialized look-up tables and reducing the number of look-up operations required.
3Reliability
If look-up tables are used for decompression, then coefficient values can be retrieved, but memory usage increases and look-up operations are required during decompression
Solution Approach 1:
The patent replaces the mechanical look-up table system with a mathematical function-based system. Instead of retrieving values from pre-computed tables stored in memory, the system evaluates parametric functions (e.g., Gaussian probability density functions) during decompression. This substitution eliminates the need for large memory structures while maintaining decompression correctness through mathematical relationships.
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.