Bridge Transform for Unsupervised Domain Generalization in Neural Networks

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing machine learning models face challenges in matching and retrieving images across different visual domains, such as real photos and schematic diagrams, without the need for manual labeling, which is costly and impractical for adapting to new technical corpora and clients.

Innovation Solution

A system that trains a neural network using a bridge transform to convert training images into a bridge domain, enabling contrastive learning and generating representations that are similar across different image domains, with a learned domain-specific model and edge maps, allowing for unsupervised domain generalization and improved performance without relying on labeled images.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manual labeling is used for training machine learning models across different visual domains, then accuracy can be improved, but time and cost increase significantly

Engineering Contradiction:

ImproveaccuracyVSAvoidtime and cost

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The system uses self-supervised learning where the model learns from unlabeled images by creating augmented versions and training through contrastive loss, eliminating the need for manual labeling while maintaining accuracy across different visual domains

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The bridge transform creates a universal bridge domain that works across multiple visual domains (real photos, sketches, diagrams, paintings), allowing a single training approach to generalize universally without domain-specific manual labeling

Inventive Principle:

Principle #6Universality (Multi-functionality)

2Measurement precision

If domain-specific models are trained for each visual domain, then accuracy improves, but device complexity increases

Engineering Contradiction:

ImproveaccuracyVSAvoidmodel complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The system merges multiple domain-specific transformations into a single bridge transform that maps different visual domains to a common bridge domain, reducing complexity while maintaining accuracy through unified contrastive learning

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The bridge domain acts as an intermediary space that connects different visual domains, allowing the model to learn domain-invariant features without requiring separate complex models for each domain

Inventive Principle:

Principle #24Intermediary (Mediator)

3Adaptability or versatility

If contrastive learning with augmented image pairs is implemented, then domain generalization improves, but computational requirements increase

Engineering Contradiction:

Improvedomain generalizationVSAvoidcomputational resources

Core Design Contradiction:

Adaptability or versatilityVSUse of energy by moving object

Solution Approach 1:

The system applies selective data augmentation techniques (cropping, flipping, color jittering) that are sufficient for learning domain-invariant features without excessive computational overhead, balancing generalization improvement with resource efficiency

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS20230306721A1Machine learning models trained for multiple visual domains using contrastive self-supervised training and bridge domain

Publication Date: 2023.09.28 INTERNATIONAL BUSINESS MACHINE CORPORATION

AI summary

An example a system includes a processor to receive a model that is a neural network and a number of training images. The processor can train the model using a bridge transform that converts the training images into a set of transformed images within a bridge domain. The model is trained using a contrastive loss to generate representations based on the transformed images.