Shared Encoder Training Across Private Multi-Domain Corpora

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

It is difficult to obtain training corpus from a plurality of fields with a large cost, making it challenging to implement pre-training models effectively.

Innovation Solution

A method for generating a shared encoder using a master node that organizes child nodes to use private training samples, determining a target parameter set based on each child node's training results, thereby sharing training corpus across multiple fields while reducing costs and improving performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If training corpus is obtained from multiple fields through traditional methods, then the performance of pre-training models is improved, but the cost and difficulty of data acquisition increase significantly

Engineering Contradiction:
Improveperformance of pre-training modelsVSAvoiddifficulty and cost of obtaining training corpus
Core Design Contradiction:
ReliabilityVSEase of manufacture

Solution Approach 1:

The patent divides the training corpus acquisition process into multiple independent fields or domains, where each field contributes its own data to a unified pre-training model. This segmentation allows organizations to leverage their existing domain-specific data without needing to acquire data from all fields centrally, reducing the difficulty and cost of data acquisition while maintaining model performance across multiple domains.

Inventive Principle:
Principle #1Segmentation

2Adaptability or versatility

If pre-training models are implemented with comprehensive multi-field data, then the model's versatility and accuracy are enhanced, but the time and resources required for data collection and processing increase

Engineering Contradiction:
Improvemodel performance across multiple fieldsVSAvoidtime and resources for data collection and processing
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent enables organizations to perform pre-training using their own existing domain-specific data before fine-tuning for specific tasks. This preliminary action eliminates the need to wait for comprehensive multi-field data to be collected centrally, allowing models to be developed and deployed more quickly while still achieving versatility through subsequent fine-tuning on task-specific data.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3855368B1Shared encoder generation method and apparatus, and electronic device
Publication Date: 2025.12.03 BEIJING BAIDU NETCOM SCI & TECH CO LTD
  • EP3855368B1 patent drawingFigure 1
  • EP3855368B1 patent drawingFigure 2
  • EP3855368B1 patent drawingFigure 3

AI summary

A method and an apparatus for generating a shared encoder, and an electronic device are provided by the present application, which belongs to a field of computer technology. The method includes: sending by a master node a shared encoder training instruction to child nodes, so that each child node obtains training samples based on a type of a target shared encoder included in the training instruction; sending an initial parameter set of the target shared encoder to be trained to each child node after obtaining a confirmation message returned by each child node, so that the initial parameter set is trained by each child node with its own training samples; obtaining an updated parameter set of the target shared encoder returned by each child node; determining a target parameter set corresponding to the target shared encoder based on a first preset rule and the updated parameter set of the target shared encoder returned by each child node. As a result, the method for generating the shared encoder may reduce the difficulty and cost of obtaining training corpus from a plurality of fields and improve the performance of the shared encoder.