Image Translation Functions for Interpreting Model Predictions

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image analysis methods, particularly in medical diagnostics, fail to identify the specific visual features that drive a machine-learned model's predictions, especially when those features are subtle or imperceptible to human observers.

Innovation Solution

Employ image translation functions, such as generative adversarial networks (GANs), to modify and exaggerate features in images, allowing for the identification of subtle or imperceptible features that influence model predictions, using translation functions to create and iteratively amplify changes in images.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of information

If traditional image analysis methods are used, then the model can make predictions, but the specific visual features driving predictions cannot be identified

Engineering Contradiction:
Improvefeature interpretabilityVSAvoidanalysis method complexity
Core Design Contradiction:
Loss of informationVSDevice complexity

Solution Approach 1:

The patent introduces image translation functions as an intermediary tool between the predictive model and human analysts. These translation functions modify input images to exaggerate subtle visual features, making them more apparent while preserving the original prediction context. This intermediary approach enables feature identification without requiring direct modification of the predictive model itself.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent applies parameter changes by systematically modifying image characteristics through translation functions. By adjusting parameters such as contrast, intensity, and spatial transformations, the system exaggerates subtle visual features to make them perceptible while maintaining the underlying predictive relationships.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If image translation functions are applied to exaggerate features, then subtle features become identifiable, but the processing complexity increases

Engineering Contradiction:
Improvefeature detection accuracyVSAvoidtranslation function complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent employs preliminary action by pre-training image translation functions on large datasets before deployment. This pre-training phase establishes the translation functions' ability to identify and exaggerate relevant visual features, so that during actual use, the complex processing is minimized to simple application of pre-learned transformations.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system incorporates feedback mechanisms where the effects of translation functions are evaluated against prediction outcomes. This feedback loop allows the system to refine which translation operations are applied and to what extent, optimizing the balance between feature exaggeration and processing complexity.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP4091132B1Explanation of machine-learned models using image translation
Publication Date: 2025.11.05 GOOGLE LLC
  • EP4091132B1 patent drawingFigure 1
  • EP4091132B1 patent drawingFigure 2
  • EP4091132B1 patent drawingFigure 3

AI summary

Systems and methods for identifying visual features that influence a predictive model are provided. The technology employs an image translation function to introduce a visual feature into an image to create a modified image that can be fed to a predictive model. When the predictive model generates a different prediction for a given image than it does for a modified version of that image, the image translation function can then be used to make further modified versions that exaggerate the introduced visual feature. The technology thus aids in identifying visual features that influence the predictive model so that the model's conclusions can be understood, and so that those visual features can be further studied and tested.