Explanatory Dropout for On-Distribution Model Reason Codes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models are complex and opaque, making it difficult for humans to understand their underlying relationships and compliance with regulatory standards, such as anti-discriminatory regulations, is challenging due to lack of explainability.
Innovation Solution
Implement explanatory dropout systems and methods that use on-manifold/on-distribution evaluation of dropout operators to train machine learning models, allowing for the calculation of numerical statistics and generation of reason codes to explain model outputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning models use complex multi-layered networks to achieve high prediction accuracy, then predictive power and accuracy are improved, but model explainability and human understanding deteriorate
Solution Approach 1:
The patent introduces dropout operators as intermediary elements that systematically remove specific features from input data during training. These operators act as mediators between the complex model and human understanding by creating controlled perturbations that reveal feature importance. The dropout operators enable explanation generation without modifying the core predictive model architecture, thus preserving accuracy while adding interpretability capabilities.
Solution Approach 2:
The patent segments the feature space into distinct groups that can be independently dropped out using dropout operators. Each operator targets specific feature groups, allowing the model to be evaluated under different feature subsets. This segmentation enables systematic analysis of feature contributions while maintaining the integrity of the overall complex model structure.
2Loss of information
If traditional off-manifold or off-distribution perturbative explanation procedures are used to explain model outputs, then some level of explanation is provided, but reliability and computational efficiency deteriorate
Solution Approach 1:
The patent applies dropout operators during the training phase rather than as post-hoc explanations. By pre-training the model with dropout operators applied to training data, the model learns to be robust to feature removals and the training process itself generates reliable explanation signals. This preliminary action ensures that explanations are based on distributions the model has actually seen during training, improving reliability.
Solution Approach 2:
The patent changes the distribution parameters by applying dropout operators that modify input data distributions in controlled ways. Instead of using arbitrary off-distribution perturbations, the dropout operators create perturbations that maintain the underlying data manifold, ensuring evaluations remain on-distribution and thus more reliable.
3Loss of information
If comprehensive feature analysis is performed to ensure regulatory compliance, then model explainability is improved, but computational complexity and processing time increase
Solution Approach 1:
The patent applies dropout operators selectively to specific feature groups rather than performing exhaustive analysis of all possible feature combinations. By focusing on targeted feature groups that are most relevant to compliance requirements, the system obtains sufficient explanation coverage without the computational burden of complete feature space exploration.
Data Source
AI summary
Explanatory dropout systems and methods for improving a computer implemented machine learning model are provided using on-manifold/on-distribution evaluation of dropout of key features to explain model outputs. The machine learning model is trained using a plurality of input examples, including input records with explicit dropout operators applied effectuating the removal of influence of features associated with an explanation reason class. One or more dropout operators may be stochastically applied to one or more input examples. The procedure includes on-manifold/on-distribution evaluation of the machine learning model under conditions of absence or presence of the one or more dropout operators for reliable calculation of numerical statistics associated with reason classes to yield model explanations. The training and evaluation procedures present advantages over traditional off-manifold or off-distribution perturbative explanation procedures.