Neural Network Input Perturbation for Efficient Explanation

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current techniques for explaining neural network operations, such as LIME, are inefficient in determining the significance of input data elements on output results, as they require processing the entire perturbed input data array, which is unnecessary and resource-intensive.

Innovation Solution

The method involves applying a perturbation to only a part of the input data array and processing only that perturbed portion through the neural network, while reusing results from the unperturbed parts, thereby reducing overall processing requirements.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of information

If perturbation-based techniques are applied to explain neural network operations, then understanding of input data significance is improved, but processing time and computational resources increase

Engineering Contradiction:

Improveunderstanding of input data significanceVSAvoidprocessing time

Core Design Contradiction:

Loss of informationVSLoss of time

Solution Approach 1:

The input data array is divided into multiple parts, and perturbations are applied to only specific parts rather than the entire array. This segmentation allows the system to focus computational resources on analyzing the impact of perturbations in specific regions, thereby reducing overall processing time while maintaining explanatory power for the neural network operations.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

Instead of applying perturbations to the entire input data array (excessive action), the method applies perturbations only to selected parts of the array (partial action). This partial perturbation approach reduces the computational burden and processing time required for explanation while still providing sufficient insight into the neural network's operation.

Inventive Principle:

Principle #16Partial or excessive action

2Loss of information

If the entire perturbed input data array is processed through the neural network, then complete explanation is obtained, but processing costs increase

Engineering Contradiction:

Improvecompleteness of explanationVSAvoidprocessing costs

Core Design Contradiction:

Loss of informationVSUse of energy by moving object

Solution Approach 1:

The input data array is segmented into multiple parts, and the method processes only the perturbed parts through the neural network while reusing results from unperturbed parts. This segmentation strategy reduces the total number of forward passes required, thereby lowering processing costs and energy consumption while maintaining sufficient explanatory completeness.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The method applies partial perturbation to only specific parts of the input data array rather than the entire array. By processing only the perturbed portions through the neural network and reusing results from unperturbed portions, the system reduces computational costs and energy consumption while still obtaining a complete explanation for the neural network's operations.

Inventive Principle:

Principle #16Partial or excessive action

3Loss of information

If perturbation is applied to the entire input data array, then comprehensive analysis is achieved, but processing efficiency decreases

Engineering Contradiction:

Improvecomprehensive analysis capabilityVSAvoidprocessing efficiency

Core Design Contradiction:

Loss of informationVSProductivity

Solution Approach 1:

The input data array is divided into multiple parts, and perturbations are applied only to selected parts. The system then processes only these perturbed parts through the neural network while reusing results from unperturbed parts. This segmentation approach maintains comprehensive analysis capability while significantly improving processing efficiency by reducing the total computational work required.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

Instead of applying perturbation to the entire input data array (excessive action), the method applies perturbation only to specific parts (partial action). This partial perturbation strategy maintains comprehensive analysis capability by focusing on the most informative regions while dramatically improving processing efficiency and reducing computational overhead.

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS20230316063A1Neural network processing

Publication Date: 2023.10.05 ARM LTD

AI summary

An input data array is subjected to neural network processing to generate a result of the neural network processing for the input data array. A perturbation is applied to a part (but not all of) the input data array, with neural network processing then performed using the so-perturbed version of the input data array. However only some (and not all) of the perturbed version is subjected to neural network processing, based on the part of the input data array to which the perturbation has been applied. The result of the neural network processing of the perturbed version of the input data array is compared with the result of the neural network processing of the input data array without the perturbation, to determine whether the perturbation of the input data array has an effect on the result of the neural network processing.