Neural Network Bias Removal Through Node Scoring and Exclusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing predictive performance models fail to detect and exclude variables that may introduce unethical or illegal biases, leading to biased outputs despite excluding protected data, as biases can emerge during recursive analysis or in hidden layers of neural networks.
Innovation Solution
A system that uses a hardware processor to receive input data, generate an artificial intelligence neural network, determine predictive bias, score nodes based on bias impact, and generate a new neural network excluding biased parameters to mitigate bias, ensuring compliance with validity and bias thresholds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If protected data is excluded from input to predictive model, then ethical compliance is improved, but bias in output may still occur due to unprotected data introducing indirect bias
Solution Approach 1:
The patent segments the predictive model into multiple components including input data, processing layers, and output layers. It specifically identifies and segments biased variables within the model structure, allowing targeted removal of bias-contributing elements while preserving the overall model functionality and ethical compliance.
Solution Approach 2:
The patent extracts and identifies biased variables from the predictive model by analyzing variable importance and relationships. It removes specific biased variables or combinations of variables that contribute to unethical outputs, thereby eliminating harm while maintaining model performance on legitimate predictive tasks.
2Measurement precision
If neural network processes data through hidden layers and recursive analysis, then predictive accuracy is improved, but biased variable combinations may emerge internally
Solution Approach 1:
The patent implements feedback mechanisms that monitor the internal processing of neural networks. It analyzes variable importance metrics and relationships between variables at different processing stages, providing feedback to identify where and how biased combinations emerge, enabling targeted intervention to prevent harmful bias generation.
Solution Approach 2:
The patent performs preliminary analysis of variable relationships and importance before final model deployment. It proactively identifies potential biased variable combinations that may emerge during processing, allowing pre-emptive removal or adjustment of problematic variables before they can generate harmful bias in outputs.
3Reliability
If initial inputs are individually unbiased, then data quality is improved, but biased outcomes may still result from variable interactions during processing
Solution Approach 1:
The patent analyzes how individual unbiased variables combine and interact during model processing. It identifies specific combinations of variables that, when merged or interacted with, generate biased outcomes. The system then adjusts or removes problematic variable combinations while preserving the individual unbiased inputs and their legitimate predictive value.
Data Source
AI summary
A hardware processor can generate an artificial intelligence neural network that is predictive of performance. The hardware processor can process the artificial intelligence neural network to determining whether a validity value for the artificial intelligence neural network meets a validity threshold. A predictive bias can be computed for the artificial neural network based on non-factored inputs. Nodes of the artificial neural network can be scored to compute an effect on the predictive bias. Another artificial intelligence neural network predictive of performance can be generated excluding a combination of parameters associated with a highest scored node of the artificial intelligence neural network.