Complex Iterated Least Square Thresholding for Neural Network Training

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing multipath filtering systems, particularly those using iterated least square thresholding (ILST) algorithms, struggle with convergence issues when dealing with complex signals and larger neural network architectures, limiting their effectiveness in modeling complex scenarios such as urban clutter and diverse radar systems.

Innovation Solution

The implementation of a complex iterated least square thresholding (CILST) algorithm, capable of processing both real and complex signals, allows for improved training of complex-valued neural networks, enabling better signal separation, parameter estimation, and signal classification in multipath environments.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If the ILST algorithm is used for non-linear parameter estimation, then real-valued processing is achieved, but complex phase information cannot be used optimally and convergence struggles with larger neural networks

Engineering Contradiction:

Improveparameter estimation accuracyVSAvoidcomplex signal processing capability

Core Design Contradiction:

Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent changes the fundamental parameter type from real-valued to complex-valued by introducing the CILST algorithm. This allows the neural network to process complex signals with phase information while maintaining the iterated least square thresholding framework, thereby resolving the contradiction between measurement precision and adaptability to complex signals

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

The patent substitutes the traditional backpropagation algorithm with the CILST algorithm for training complex-valued neural networks. This replacement eliminates convergence problems associated with backpropagation in complex domains and enables optimal use of complex phase information, improving both precision and versatility

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

2Device complexity

If the neural network has more than a small number of hidden layer neurons, then modeling capability increases, but the ILST algorithm struggles with convergence

Engineering Contradiction:

Improveneural network architecture complexityVSAvoidalgorithm convergence

Core Design Contradiction:

Device complexityVSReliability

Solution Approach 1:

The patent replaces the ILST training algorithm with CILST, which is specifically designed to handle complex-valued neural networks. This substitution enables reliable convergence even with larger numbers of hidden layer neurons, allowing the system to maintain both high modeling complexity and algorithmic reliability

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

By changing from real-valued to complex-valued parameter processing in the training algorithm, the patent enables stable convergence with larger neural network architectures. The CILST algorithm's complex arithmetic capabilities allow it to properly handle the additional degrees of freedom introduced by more neurons without suffering from the convergence issues that plague real-valued algorithms

Inventive Principle:

Principle #35Parameter changes

3Productivity

If backpropagation is used for training, then neural network training is achieved, but convergence problems occur especially for larger networks

Engineering Contradiction:

Improvetraining capabilityVSAvoidconvergence stability

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The patent substitutes the backpropagation algorithm with CILST for training neural networks. This replacement eliminates the convergence problems that plague backpropagation, especially in larger networks, while maintaining the ability to train complex-valued networks. The CILST algorithm provides both productivity and reliability by ensuring stable convergence

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent introduces CILST as an intermediary training algorithm that bridges the gap between the need for efficient neural network training and the requirement for stable convergence. Rather than directly using backpropagation, the CILST algorithm serves as a mediator that transforms the training process to achieve both speed and reliability

Inventive Principle:

Principle #24Intermediary (Mediator)

4Measurement precision

If global optimization is used before backpropagation, then starting point accuracy improves, but convergence problems persist and can lead to inaccurate starting points

Engineering Contradiction:

Improvestarting point accuracyVSAvoidconvergence stability

Core Design Contradiction:

Measurement precisionVSReliability

Solution Approach 1:

The patent extracts and eliminates the problematic two-stage training process (global optimization followed by backpropagation) and replaces it with a single CILST algorithm. This removal of the intermediate global optimization step eliminates the source of convergence problems while maintaining training effectiveness, achieving both accuracy and reliability in one unified process

Inventive Principle:

Principle #2Taking out (Extraction)

Data Source

PatentUS10275707B2Systems and methods for training multipath filtering systems

Publication Date: 2019.04.30 THE BOEING CO

AI summary

A method for training a neural network to be configured to filter a multipath corrupted signal is provided. The method includes receiving, at the neural network, real or simulated multipath corrupted signal data, and training the neural network on the multipath corrupted signal data using a complex iterated least square thresholding algorithm (CILST) capable of processing both real and complex signals.