Neural Network Negative Rule Extraction via Input Perturbation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods are ineffective in identifying rules, especially negative rules, in complex neural networks with multiple hidden layers, leading to instability and computational inefficiency, and lack the ability to account for absent features in input data.
Innovation Solution
A controller is used to identify negative rules by applying perturbations to training datasets, analyzing the minimal changes that cause different classifications, and focusing on missing features to extract rules that account for the absence of elements in inputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If decompositional techniques are used to identify rules by examining hidden layers, then rules can be identified for simple neural networks, but these techniques fail substantially for complex neural networks with more than two hidden layers
Solution Approach 1:
The patent introduces an intermediary system that observes the neural network's input-output behavior without examining the internal hidden layers. This intermediary approach allows rule extraction for complex networks by treating the network as a black box, avoiding the complexity problem while maintaining rule identification capability through observational analysis of input-output relationships
2Measurement precision
If pedagogical techniques are used to identify rules by treating the neural network as a black box, then rules can be identified for stable input sets, but these techniques become computationally expensive and produce unstable contradictory rules when inputs change
Solution Approach 1:
The patent extracts only the essential input-output relationships needed for rule identification, removing unnecessary computational overhead. By focusing solely on observing how inputs map to outputs without comprehensive black-box analysis, the system achieves rule extraction with reduced computational cost and improved stability when inputs change
Solution Approach 2:
The system uses feedback from the neural network's classifications to refine rule identification. By observing how the network responds to different inputs and using this feedback to adjust the extracted rules, the system maintains computational efficiency while achieving stable rule sets even when inputs change notably
3Loss of information
If traditional rule identification methods are used, then positive rules can be identified, but the ability to identify negative rules (absent features) is lacking
Solution Approach 1:
The patent inverts the traditional approach by focusing on what is absent rather than what is present. Instead of identifying rules based on present features, the system identifies negative rules by observing how the neural network responds when features are missing or removed from inputs, thereby capturing the complete picture of feature relationships including absent features
Data Source
AI summary
A first classification is received from a neural network regarding a training dataset sent to the neural network. A modified training dataset with a perturbation of the training dataset is identified, where this modified training dataset causes the neural network to return a second classification. The perturbation is analyzed to identify a negative rule of the neural network.