Computing Relative Impact of Training Examples for ML Explainability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models lack explainability, making it difficult to identify biases and debug or improve them, especially when trained on large datasets with mislabeled information, which is crucial for making accurate decisions and complying with regulations like GDPR.
Innovation Solution
A system that computes relative numerical impact values for training examples, allowing identification of the most influential examples and mislabeled data, enabling adjustments to the model without removing training datasets, and providing explanations for model outputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning models use large amounts of training data to generate accurate outputs, then model accuracy is improved, but the ability to explain and debug the model decreases
Solution Approach 1:
The patent segments the training data into individual training examples and calculates the relative numerical impact of each example on the model output. This segmentation allows the system to attribute specific outputs to specific training data points, enabling explainability while maintaining the benefits of using large datasets for training.
Solution Approach 2:
The patent introduces an intermediary mechanism that computes and stores the relative numerical impact values between training examples and model outputs. This intermediary layer bridges the gap between the black box model and human users, providing explanations without compromising the model's accuracy or the size of the training data.
2Productivity
If machine learning models operate as black boxes with hidden learned relationships, then computational efficiency is improved, but the ability to identify and fix biases decreases
Solution Approach 1:
The patent performs preliminary computation of relative numerical impact values for all training examples during or after training. This preliminary action stores the necessary information to explain model outputs and identify biases later, without requiring additional computational resources during inference or compromising the black box efficiency.
Solution Approach 2:
The patent implements a feedback mechanism where the relative numerical impact values are used to identify which training examples contribute most to specific outputs. This feedback loop enables users to review and correct biased training data, improving model reliability while maintaining computational efficiency during normal operation.
3Adaptability or versatility
If machine learning models are trained on diverse data sources including crowd-sourced data, then model versatility is improved, but the difficulty of identifying mislabeled data increases
Solution Approach 1:
The patent segments the diverse training data into individual examples from different sources and calculates their specific impact on model outputs. This segmentation makes it possible to trace and identify mislabeled data points regardless of their source, enabling quality control while maintaining the versatility gained from diverse data sources.
Solution Approach 2:
The patent replaces manual inspection of training data with an automated computational mechanism that calculates relative numerical impacts. This substitution eliminates the need for manual review of each training example, making mislabeled data identification scalable across diverse and large datasets.
4Reliability
If machine learning models provide detailed explanations for outputs, then regulatory compliance is improved, but the complexity of the system increases
Solution Approach 1:
The patent performs the complex computations of relative numerical impact values during training or as a preliminary step, rather than during each explanation request. This preliminary action shifts the computational complexity to a batch process, making the explanation generation process simple and compliant with regulatory requirements.
Solution Approach 2:
The patent creates a simplified representation of the training data's impact on model outputs through relative numerical impact values. This copying approach provides the necessary information for regulatory compliance without requiring access to or manipulation of the full complex training datasets or model parameters.
Data Source
AI summary
Systems and methods for improving a machine learning model are described. In an embodiment, a computing system stores a plurality of training examples comprising training inputs and training outputs. The computing system generates a machine learning model and training the machine learning model using the plurality of training examples. The computing system receives a particular input for the machine learning system and, using the particular input and the machine learning system, computes a particular output. For each training example of the plurality of training examples, the computing system adjusts a weight of the training example on the machine learning system and computes a relative numerical impact on the particular output for the training example, the relative numerical impact reflecting an importance of each training example on the particular output relative to an importance of the other training examples of the plurality of training examples on the particular output. The server computer generates training example relevance data comprising identifiers of the plurality of training examples and the relative numerical impact values for the plurality of training examples. The server computer stores the training example relevance data in the computing system.