Generative Model Output Attribution via Latent Similarity

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing approaches to attributing generative model outputs to training data samples are complex and require multiple rounds of computationally intensive retraining and fine-tuning, making them inefficient for determining the contribution of content items to the output of a generative model.

Innovation Solution

A technique that involves determining latent representations of a generative output and associated data samples, computing similarities between these representations, and generating output indicating high similarity between the generative output and specific data samples, thereby attributing characteristics of the generative output to individual data samples.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If existing approaches (ensemble training with ablation or fine-tuning) are used to attribute generative output to training data samples, then attribution accuracy can be achieved, but computational complexity and time consumption increase significantly

Engineering Contradiction:
Improveattribution accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts only the necessary components for attribution by using a single pre-trained generative model and comparing its output directly against training data samples through similarity metrics, eliminating the need for complex ensemble training and ablation procedures while maintaining attribution accuracy

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent performs preliminary action by pre-training the generative model once on the complete training dataset before attribution, then using this pre-trained model for efficient similarity-based attribution without requiring repeated retraining or fine-tuning rounds

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If existing approaches (ensemble training with ablation or fine-tuning) are used to attribute generative output to training data samples, then attribution accuracy can be achieved, but processing time increases significantly

Engineering Contradiction:
Improveattribution accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent extracts only the essential attribution function by using direct similarity comparison between generative output and training samples, removing the time-consuming iterative retraining and fine-tuning processes while preserving attribution accuracy

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent performs preliminary action by completing model training beforehand, allowing subsequent attribution operations to proceed efficiently through simple similarity computations without repeated training cycles

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If existing approaches (ensemble training with ablation or fine-tuning) are used to attribute generative output to training data samples, then attribution accuracy can be achieved, but resource overhead increases significantly

Engineering Contradiction:
Improveattribution accuracyVSAvoidcomputational resource overhead
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent extracts only the minimal necessary computational resources by using a single pre-trained model and direct similarity metrics, eliminating the excessive resource consumption associated with ensemble training and iterative fine-tuning while maintaining attribution accuracy

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP4495831A1Attribution of generative model outputs
Publication Date: 2025.01.22 DISNEY ENTERPRISES INC
  • EP4495831A1 patent drawingFigure 1
  • EP4495831A1 patent drawingFigure 2
  • EP4495831A1 patent drawingFigure 3

AI summary

The present invention sets forth a technique for analyzing a generative output of a generative model. The technique includes determining a first latent representation of the generative output and a plurality of latent representations of a plurality of data samples associated with the generative model. The technique also includes computing a plurality of similarities between the first latent representation and the plurality of latent representations. In response to determining that a first similarity that is included in the plurality of similarities and computed between the first latent representation and a second latent representation included in the plurality of latent representations exceeds a threshold, the technique includes causing output to be generated that indicates a high similarity between the generative output and a first data sample that is included in the plurality of data samples and corresponds to the second latent representation.