Multi-Task Latent Space Alignment for Small Molecular Data

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing transfer learning techniques struggle with small, complex data sets such as molecular structure data, particularly in non-Euclidean spaces, leading to inadequate performance and stability in high-dimensional problems.

Innovation Solution

A multi-tasking model training method that utilizes geometric alignment in an integrated latent space to transfer and learn knowledge across tasks, incorporating geometric alignment vectors, perturbation vectors, and multiple loss functions to optimize model performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If existing Euclidean space-based transfer learning techniques are used, then the model can be applied to new domains, but the prediction performance and stability are inadequate on small, complex data sets like molecular structure data

Engineering Contradiction:
Improveprediction performance and stabilityVSAvoidapplicability to new domains and tasks
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The patent changes the fundamental parameter of the feature space from Euclidean to Riemannian manifold, allowing the model to better capture complex relationships in molecular structure data while maintaining transfer learning capabilities across domains

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent introduces curved manifold geometry to represent latent spaces, where the curvature captures the intrinsic structure of molecular data. This geometric transformation enables better generalization to new domains by preserving structural relationships that flat Euclidean space cannot represent

Inventive Principle:
Principle #14Spheroidality (Curvature)

2Quantity of substance

If transfer learning is applied to small data sets, then the need for large amounts of data is reduced, but the handling of complex structures in non-Euclidean spaces becomes inadequate

Engineering Contradiction:
Improveamount of data requiredVSAvoidhandling of complex structures
Core Design Contradiction:
Quantity of substanceVSDifficulty of detecting and measuring

Solution Approach 1:

The patent transforms the data representation parameter from Euclidean vectors to Riemannian manifold points, enabling effective transfer learning on small data sets by capturing complex molecular structures through geometric properties rather than requiring large volumes of training data

Inventive Principle:
Principle #35Parameter changes

3Adaptability or versatility

If geometric alignment in integrated latent space is implemented, then knowledge transfer between tasks is improved, but the model complexity increases

Engineering Contradiction:
Improveknowledge transfer across tasksVSAvoidmodel complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent creates a universal Riemannian manifold framework that serves multiple tasks simultaneously. The geometric alignment mechanism provides a unified approach for knowledge transfer across different molecular properties and domains, reducing the need for task-specific model components despite the increased geometric complexity

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP4679330A1Multi-tasking model training method and multi-tasking performing method using machine learning model trained on basis thereof
Publication Date: 2026.01.14 LG MANAGEMENT DEV INST CO LTD
  • EP4679330A1 patent drawingFigure 1
  • EP4679330A1 patent drawingFigure 2
  • EP4679330A1 patent drawingFigure 3

AI summary

A multi-tasking model training method and a multi-tasking performing method using a machine learning model trained on the basis thereof, according to an embodiment of the present invention, may mutually transfer and learn knowledge data of a latent space for each task through geometric alignment in one integrated latent space in order to process a multi-task for output according to a plurality of domains.