Multi-source Transfer Learning via Maximal Correlation Weighting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional multi-task learning techniques require control over source datasets or access to training data, which is impractical or unethical in cases where privacy or data availability is limited, especially when using pre-trained networks for new classification tasks with few samples.
Innovation Solution
The implementation of maximal correlation weighting (MCW) for combining pre-trained neural networks to perform few-shot learning by computing maximal correlation functions between feature functions and target task samples, allowing for decoupled weighting of feature functions to build classifiers without access to source task samples.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional multi-task learning techniques are used to train neural networks on multiple source datasets, then the classification performance can be improved, but it requires control over source datasets or access to training data which is impractical when privacy or data availability is limited
Solution Approach 1:
The patent extracts only the necessary feature functions from pre-trained neural networks without requiring access to the original training data. By taking out the feature extraction capability from the trained models, the system enables transfer learning while preserving data privacy and eliminating the need for source dataset control.
Solution Approach 2:
The patent introduces an intermediary mechanism that computes maximal correlation functions between feature functions and target task samples. This intermediary computation enables the system to leverage pre-trained networks for few-shot learning without direct access to source training data, bridging the gap between model capabilities and data privacy requirements.
2Speed
If pre-trained networks are used for few-shot learning on new classification tasks, then the adaptation speed is improved, but the ability to combine multiple source networks is limited by conventional techniques
Solution Approach 1:
The patent creates a universal framework that can combine multiple pre-trained neural networks with different architectures and training histories. The maximal correlation weighting mechanism serves as a multi-functional tool that adapts to various source networks and target tasks, enabling flexible combination of diverse pre-trained models for few-shot learning.
Solution Approach 2:
The patent changes the parameter weighting scheme by computing maximal correlation functions between feature functions and target task samples. This parameter transformation enables dynamic adjustment of the contribution from each pre-trained network based on its relevance to the target task, optimizing the combination for rapid adaptation.
3Object-affected harmful factors
If access to source task samples is restricted for privacy reasons, then data privacy is protected, but the ability to train effective classifiers on target tasks is reduced
Solution Approach 1:
The patent segments the learning process into two independent phases: (1) training pre-trained networks on source datasets with full data access, and (2) applying these networks to target tasks with only feature function access. This segmentation isolates the privacy-sensitive training phase from the privacy-preserving application phase, maintaining both classifier effectiveness and data privacy.
Solution Approach 2:
The patent creates copies of the source training data representations through pre-trained neural networks. Instead of accessing the original source samples, the system uses the learned feature functions as copies that capture the essential patterns, enabling effective target task training while preserving the privacy of the original data.
Data Source
AI summary
Examples described herein provide a computer-implemented method that includes training, by one or more processing devices, a first neural network for classification based on training data in accordance with a first learning objective, the first neural network producing an intermediate feature function and a final feature function as outputs. The computer-implemented method further includes training, by the one or more processing devices, a second neural network for classification based on the intermediate feature function and the final feature function and further based at least in part on target task samples in accordance with a second learning objective. Training the second neural network includes computing maximal correlation functions of each of the intermediate feature function, the final feature function, and the target task samples.