ML Model Logic Alignment via Explanation Differentiation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
There is a challenge in determining the quality and accuracy of machine learning model outputs, which can be addressed by aligning the logic of a machine learning model with its associated explanation to improve output quality and accuracy.
Innovation Solution
A method involving a processor that receives raw data, trains a model, determines background data, computes explanations using SHAP techniques, and calculates a distance metric to align the model logic with its explanation, identifying important and non-important features to refine the model outputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning models are trained using historical data to perform AI tasks, then the model can generate outputs for various tasks, but the quality and accuracy of model outputs cannot be properly determined without a method to align model logic with explanations
Solution Approach 1:
The patent introduces an intermediary alignment mechanism that bridges model logic and explanations. The system generates explanations for model predictions, compares these explanations with the actual model logic, and uses this comparison to determine output quality and accuracy. This intermediary alignment process resolves the contradiction by providing a systematic way to measure model quality without directly complicating the model structure itself.
Solution Approach 2:
The patent replaces traditional mechanical validation methods with an explanation-based alignment approach. Instead of using complex computational mechanisms to directly verify model outputs, the system substitutes this with a semantic alignment process between explanations and model logic, using natural language processing and comparison techniques to determine accuracy and quality.
2Measurement precision
If model logic is aligned with explanations to improve output quality, then accuracy determination becomes possible, but the process requires computing multiple differentiators and distance metrics which increases computational complexity
Solution Approach 1:
The patent segments the alignment process into distinct computational components: generating explanations, extracting model logic, computing differentiators, calculating distance metrics, and determining alignment scores. By dividing the overall alignment task into these manageable segments, the system reduces computational complexity while maintaining accurate output determination. Each segment can be processed independently and efficiently.
Solution Approach 2:
The patent applies partial action by computing only the necessary differentiators and distance metrics required for alignment determination, rather than performing exhaustive analysis of all model parameters. The system focuses on computing explanations and their alignment with model logic to the extent needed for accurate quality assessment, avoiding unnecessary computational overhead.
3Loss of information
If explanations are computed based on model outputs and background data, then feature importance information is obtained, but the initial explanations may not accurately reflect the true model logic requiring iterative refinement
Solution Approach 1:
The patent implements a feedback mechanism where initial explanations are generated, compared with actual model logic, and the alignment assessment feeds back into refining the explanations. The system computes distance metrics between explanations and model logic, then uses this feedback to iteratively improve explanation accuracy. This feedback loop ensures feature importance information becomes increasingly accurate while managing refinement time through controlled iteration.
Solution Approach 2:
The patent performs preliminary action by generating initial explanations before full alignment verification. These preliminary explanations provide immediate feature importance information that can be used preliminarily, while the system simultaneously or subsequently refines them through alignment comparison. This allows the system to provide useful information quickly while still achieving accurate alignment through iterative refinement.
Data Source
AI summary
Methods and systems for aligning a logic of a machine learning model with an explanation associated with the model are provided. The method includes: receiving raw data and training a model by using the raw data; determining common background data based on the raw data; determining an inference that indicates a logic of the model; computing a first explanation about the inference based on an output of the model and the common background data; computing a model differentiator that indicates an item that is important to the logic but not to the first explanation; computing an explanation differentiator that indicates an item that is important to the first explanation but not to the logic; computing a distance metric that indicates a degree of similarity between the logic and the first explanation; and computing a second explanation that includes a respective ranking score for each feature that affects the inference.