Neural Network Selection Using Unsupervised Disentanglement Similarity

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing neural networks struggle to generate disentangled representations without supervision, relying on supervised knowledge for disentanglement metrics, which limits the effectiveness of hyperparameter search and model selection.

Innovation Solution

A method for unsupervised identification of neural networks that generate disentangled latent variable representations by performing pairwise similarity measurements between trained networks, using techniques like Spearman's rank correlation and Lasso regression to determine a disentanglement score, allowing for parallel processing on distributed systems.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If supervised knowledge is used for disentanglement metrics, then measurement precision of disentanglement quality is improved, but ease of operation and applicability to unsupervised learning deteriorates

Engineering Contradiction:
Improvedisentanglement quality measurementVSAvoidunsupervised model selection
Core Design Contradiction:
Measurement precisionVSEase of operation

Solution Approach 1:

The patent introduces an intermediary similarity measurement mechanism that compares latent representations across multiple trained neural networks without requiring ground truth labels. This intermediary approach (pairwise similarity measurement) bridges the gap between having trained models and being able to evaluate their disentanglement quality, enabling unsupervised selection of models with high disentanglement properties through metrics like Spearman's rank correlation and Lasso regression

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The system enables neural networks to self-evaluate their disentanglement quality through intrinsic properties of their latent representations. By computing pairwise similarities between latent codes from different networks and analyzing consistency patterns, the method allows models to automatically identify which networks produce disentangled representations without external supervision or ground truth

Inventive Principle:
Principle #25Self-service

2Adaptability or versatility

If multiple neural networks are trained with different hyperparameters, then adaptability and model selection capability are improved, but device complexity and computational requirements increase

Engineering Contradiction:
Improvehyperparameter search capabilityVSAvoidcomputational system complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent segments the complex task of evaluating disentanglement quality into manageable pairwise comparisons between neural networks. Instead of computing a global disentanglement metric that would be computationally intensive, the method divides the evaluation into individual pairwise similarity measurements between latent representations, which can then be aggregated to determine overall disentanglement scores for model selection

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system evaluates neural networks trained with different hyperparameter configurations by measuring changes in their latent representation similarities. By parameterizing the evaluation around hyperparameter variations and their impact on latent space consistency, the method enables efficient hyperparameter search through identifying which parameter settings produce networks with consistent latent representations across the pool

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If pairwise similarity measurements are performed between all trained networks, then measurement precision of disentanglement score is improved, but loss of time and computational efficiency worsens

Engineering Contradiction:
Improvedisentanglement score accuracyVSAvoidcomputation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent performs preliminary encoding of all training data items through each neural network in the pool before the actual similarity measurement phase. This preliminary action prepares the latent representations in advance, allowing the subsequent pairwise similarity computations to proceed efficiently without re-processing the input data, thus reducing overall computation time while maintaining measurement precision

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3903239B1Identifying neural networks that generate disentangled representations
Publication Date: 2026.03.04 GDM HOLDING LLC
  • EP3903239B1 patent drawingFigure 1
  • EP3903239B1 patent drawingFigure 2
  • EP3903239B1 patent drawingFigure 3(1)~3(4)

AI summary

A method for automatically identifying a computer-implemented neural network which is able to generate a disentangled latent variable representation of an input data item. The method involves obtaining a pool of trained neural networks, encoding a set of evaluation data items using each of the trained neural networks to determine a respective set of latent representations for each of the trained neural networks, and determining a measure of similarity between the sets of latent representations in order to select a trained neural network with a disentangled latent variable representation.