Confidence Measure for Deployed Machine Learning Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deployed machine learning models lack a confidence measure specific to input data, making it difficult to assess the reliability of their outputs, particularly in medical applications where precise accuracy is crucial.
Innovation Solution
A method is proposed to determine a confidence measure by processing modified instances of input data through the ML model, analyzing the variance of results to assess the model's robustness and providing a data-specific confidence measure, which can be visualized with overlays and textual descriptions to aid experts in evaluating model outputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a deployed ML model is used to process input data, then a result is generated, but no confidence measure specific to the input data is provided
Solution Approach 1:
The patent introduces an intermediary confidence measurement system that processes the relationship between input data and model outputs. Instead of modifying the core ML model, a separate confidence measurement module is introduced that takes the input data and model result as inputs and generates a confidence score. This intermediary component resolves the contradiction by providing measurement precision without requiring changes to the deployed model's internal structure.
Solution Approach 2:
The patent replaces the need for complex model structure modifications with a computational algorithmic approach. Instead of physically or structurally changing the ML model to provide confidence measures, the system uses mathematical computations on the model's outputs and input data to derive confidence scores. This substitution of mechanical/structural changes with computational methods resolves the contradiction between measurement precision and device complexity.
2Reliability
If the ML model is sealed for client-side deployment, then licensing and security requirements are met, but client-side retraining and model adjustment become impossible
Solution Approach 1:
The patent enables the client-side system to self-assess the reliability of model outputs without requiring access to the model's internal parameters or retraining capabilities. The confidence measurement system operates autonomously using only the sealed model's outputs and the original input data, allowing the client to evaluate and potentially filter results based on confidence thresholds. This self-service approach maintains model sealing while providing adaptability in result interpretation.
Solution Approach 2:
The patent introduces a feedback mechanism where confidence measures are generated based on the relationship between input data and model outputs. This feedback loop allows the system to assess model performance on specific inputs without requiring model retraining. The confidence information can be used to adjust decision-making processes, request human review for low-confidence predictions, or trigger data collection for future model improvements, thereby providing adaptability within the constraints of a sealed model.
3Quantity of substance
If augmentation methods are used during training, then training dataset size and variability increase, but the model still lacks input-specific confidence measures after deployment
Solution Approach 1:
The patent applies preliminary action by generating multiple modified versions of the input data (similar to augmentation techniques) before processing them through the deployed model. This pre-processing step creates a set of variant inputs that are then used to assess model confidence. By performing this augmentation-like operation at inference time rather than training time, the system obtains input-specific confidence measures without requiring additional training data or model retraining.
Solution Approach 2:
The patent introduces dynamics by making the confidence measurement process adaptive to each specific input data instance. Rather than providing a fixed or global confidence level for the entire model, the system dynamically generates confidence measures tailored to each input by processing it through multiple augmented variants. This dynamic approach ensures that confidence measures are specific to each prediction, resolving the contradiction between training data quantity and output measurement precision.
Data Source
AI summary
Presented are concepts for obtaining a confidence measure for a machine learning model. One such concept process input data with the machine learning model to generate a primary result. It also generate a plurality of modified instances of the input data and processes the plurality of modified instances of the input data with the machine learning model to generate a respective plurality of secondary results. A confidence measure relating to the primary result is determined based on the secondary results.