Multi-Task Large Model Training with Universal and Dedicated Adapters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large models face interference and negative transfer when fine-tuned for multiple task types, leading to reduced effectiveness in specific tasks due to the 'seesaw effect' and heterogeneity between training and downstream data sets.
Innovation Solution
Implement a bypass task network with universal and dedicated adapters to process embedding vectors, performing weighted summation to maintain the original model parameters, reducing the seesaw effect and enhancing generalization across tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the large model is fine-tuned simultaneously for multiple task types, then the model can be applied to various tasks, but task signals from different task types interfere with each other causing the 'seesaw effect' where capability for one task enhances while capability for another task reduces
Solution Approach 1:
The patent segments the fine-tuning process by introducing separate adapter modules for different task types. Each adapter is dedicated to processing specific task signals, preventing interference between tasks. The adapter layer is divided into task-specific adapters that process different task types independently, while sharing common embedding layers, thus achieving multi-task applicability without the seesaw effect.
Solution Approach 2:
The patent introduces adapter modules as intermediary components between the embedding layer and task-specific processing layers. These adapters act as mediators that transform and process task signals independently, allowing different task types to be processed simultaneously without direct interference. The adapter layer serves as an intermediary that decouples task-specific processing from the main model architecture.
2Adaptability or versatility
If direct fine-tuning is performed on the large model for downstream tasks, then the model can be adapted to specific tasks, but negative transfer occurs where the fine-tuned model becomes less effective than the pre-trained model due to heterogeneity between datasets
Solution Approach 1:
The patent segments the model architecture into pre-trained base models and task-specific adapter modules. The adapter modules are added as separate components that can be trained independently on downstream task data without modifying the pre-trained base model. This segmentation allows the model to adapt to downstream tasks while preserving the effectiveness of the pre-trained parameters, preventing negative transfer.
Solution Approach 2:
The patent creates copyable adapter modules that can be independently trained and applied to different downstream tasks. Instead of directly fine-tuning the entire large model, the method creates separate adapter copies that inherit the pre-trained model's architecture and parameters, allowing task-specific adaptation without altering the original pre-trained model. This copying approach preserves model effectiveness while enabling adaptation.
3Adaptability or versatility
If the bypass task network uses multiple adapters with different task types, then the model can handle diverse tasks, but the complexity of managing multiple adapters and weight parameters increases
Solution Approach 1:
The patent implements a universal adapter architecture where a single adapter module can handle multiple task types through shared embedding layers and parameter efficient design. The adapter layer uses task-type embedding vectors that can be dynamically selected based on the input task, allowing one adapter structure to serve multiple functions. This universality reduces the need for completely separate adapter modules for each task type, thereby reducing management complexity.
Solution Approach 2:
The patent merges common processing functions into shared components, particularly in the embedding layer and adapter structure. Multiple adapters share common embedding layers and parameter structures, combining their functionality while maintaining task-specific capabilities. This merging approach reduces the total number of independent parameters and simplifies the management of multiple adapters, as they can be handled through a unified architectural framework.
Data Source
AI summary
Embodiments of this specification relate to a multi-task large model training method and apparatus. The method includes: obtaining a first embedding vector corresponding to a first sample, where the first sample has a first task type; separately inputting the first embedding vector into the target network layer to perform target processing, and inputting the first embedding vector into the bypass task network to perform bypass processing, where the bypass processing includes: separately processing the first embedding vector by using the several universal adapters and a first dedicated adapter corresponding to the first task type, and performing weighted summation on processing results of the adapters, to obtain a second embedding vector; determining a prediction result based on the second embedding vector and a third embedding vector output through the target processing; and updating the bypass task network based on a loss corresponding to the prediction result.