Cross-Modal Knowledge Transfer Using Batch-Normalization Statistics

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for cross-modal knowledge transfer require access to task-relevant paired data or source data, which may not be feasible due to memory or privacy concerns, especially when adapting from one modality to another without labeled data.

Innovation Solution

The SOCKET framework uses task-irrelevant paired datasets and batch normalization statistics to bridge the modality gap by matching feature distributions, employing modality-specific and agnostic loss functions to train target models without access to source data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If cross-modal knowledge transfer is performed using existing methods, then knowledge can be transferred from source to target modality, but access to task-relevant source data is required which may not be feasible due to memory or privacy concerns

Engineering Contradiction:
Improveknowledge transfer effectivenessVSAvoiddata accessibility
Core Design Contradiction:
ReliabilityVSEase of operation

Solution Approach 1:

The patent extracts and utilizes batch normalization statistics (mean and variance) from the source model as a surrogate for the unavailable source data. This extraction principle allows the method to operate without direct access to source data while still capturing essential statistical properties needed for effective cross-modal knowledge transfer.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent introduces task-irrelevant paired data as an intermediary between source and target modalities. This intermediary data serves as a bridge to align feature distributions across modalities without requiring access to task-relevant source data, effectively mediating the knowledge transfer process under data accessibility constraints.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If task-irrelevant paired data is used to bridge the modality gap, then the modality gap is reduced and performance improves, but additional data processing and computation are required

Engineering Contradiction:
Improveclassification performanceVSAvoidprocessing complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent performs preliminary alignment of feature distributions using task-irrelevant paired data before the actual knowledge transfer process. By pre-aligning the statistical properties of features across modalities through batch normalization statistics matching, the method simplifies subsequent training and improves overall efficiency despite the additional initial processing step.

Inventive Principle:
Principle #10Preliminary action

3Reliability

If batch normalization statistics are matched to reduce modality gap, then feature distribution alignment is achieved, but computation and storage of statistics are required

Engineering Contradiction:
Improvefeature distribution alignmentVSAvoidcomputational resources
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent utilizes batch normalization statistics which are already computed and stored during source model training. These statistics serve as a cheap, readily available surrogate for source data that requires minimal additional computational resources to extract and utilize, making the approach efficient despite the need for statistical matching.

Inventive Principle:
Principle #27Cheap short-living objects (Disposable)

Data Source

PatentEP4441661B1System and method for cross-modal knowledge transfer without task-relevant source data
Publication Date: 2025.10.29 MITSUBISHI ELECTRIC CORP
  • EP4441661B1 patent drawingFigure 1
  • EP4441661B1 patent drawingFigure 2
  • EP4441661B1 patent drawingFigure 3

AI summary

A cross-modality knowledge transfer system is provided for adapting one or more source model networks to one or more target model networks. The system is configured to perform steps of providing the TI paired datasets through the source feature encoders of the one or more source model networks, extracting TI source features and TI source moments from the TI paired data by the BN layers of the one or more source model networks, providing the TI paired datasets and the unlabeled TR datasets through the one or more target model networks to extract TI target features and TR target moments, training jointly all the feature encoders of the one or more target model networks by matching the extracted TI target features and TR target moments with the TI source features and TI source moments along with mixing weights, and forming a final target model network by combining the trained one or more target model networks.