Subspace-Constrained Mode Estimation for Low-Complexity Weight Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current partial-update methods (PUMs) for adaptive signal processors are limited by severe misadjustment and can only be used with a small number of optimization methods, particularly those with linear complexity, and fail to effectively reduce adapt-path complexity in high-dimensional processing.
Innovation Solution
The method involves performing a linear transformation of processor parameters from M-dimensions to (M1+L)-dimensions, where M1 weights are updated without constraints and M0=M−M1 weights are subjected to L soft constraints, allowing for reduced-dimensionality weights to be adapted using the same optimization strategy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If partial-update methods are used to reduce adapt-path complexity, then the number of operations is reduced, but severe misadjustment occurs and applicability is limited to linear complexity methods
Solution Approach 1:
The patent transforms the weight update problem from the original M-dimensional space to a reduced (M1+L)-dimensional subspace by applying a linear transformation matrix. This dimensionality reduction allows partial-update methods to be applied more effectively while maintaining better convergence properties and reducing misadjustment through the structured subspace constraint.
Solution Approach 2:
The patent changes the parameter representation by introducing a transformation matrix that maps full-dimension weights to reduced-dimension weights. This parameter transformation enables the adaptive processor to operate with fewer update operations while maintaining accuracy through the structured relationship between original and transformed parameters.
2Quantity of substance
If partial-update methods are used, then update set size is reduced, but the methods can only be applied to a small number of optimization methods
Solution Approach 1:
The patent creates a universal framework that can be applied to multiple optimization methods (LMS, NLMS, RLS, affine projection) by introducing a transformation matrix that works with any optimization algorithm. This makes the partial-update approach versatile and applicable beyond just linear complexity methods, as the transformation is algorithm-agnostic.
Solution Approach 2:
By transforming the weight update problem into a different parameter space through the linear transformation matrix, the patent enables partial-update methods to be applied universally across different optimization algorithms. The transformation preserves the essential structure needed for various optimization methods while reducing the update set size.
3Measurement precision
If full-dimension weight updates are performed, then accuracy is maintained, but computational complexity and cost increase significantly
Solution Approach 1:
The patent extracts only the essential degrees of freedom needed for accurate weight updates by transforming to a reduced-dimension subspace. Instead of updating all M weights, the method updates only M1+L transformed weights, extracting the critical information while discarding redundant dimensions, thus reducing computational complexity while maintaining accuracy.
Solution Approach 2:
The patent changes the dimensionality of the weight update problem from M dimensions to (M1+L) dimensions through a linear transformation. This dimensionality reduction maintains the essential accuracy requirements by preserving the signal subspace while eliminating redundant dimensions, resulting in lower computational complexity for the same level of accuracy.
Data Source
AI summary
A method is described for reduced-complexity estimation of signal and data modes in high-dimensional data sets, by implementing subspace-constrained partial updates to optimize an eigenvalue-based objective function. The method selects, from a set of combiner weights, a set of update weights and a set of held weights; performs updates to the set of held weights within a reduced-dimensionality subspace and unconstrained updates to the set of update weights to produce updated combiner weights; and employs the updated combiner weights to determine at least one solution to an eigenequation or pseudo-eigenequation.