Transfer Learning Model Training Without Over-Constraining Parameters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing transfer learning methods in machine learning impose excessive constraints on parameters, leading to undesirable accuracy and reliability of target models trained on new tasks, particularly when dealing with limited sample sizes.
Innovation Solution
The method involves constructing a target model based on the number of classes for a target task, obtaining values for regularized and cross-entropy loss functions, and updating parameters using gradient descent to avoid excessive constraints, thereby enhancing the training effect of the source model on the target task.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If parameters are constrained during transfer learning (as in FT and SPAR methods), then over-fitting on target task is prevented, but the training effect of source model on new task is damaged and accuracy/reliability deteriorates
Solution Approach 1:
The patent changes the parameter being constrained from the model parameters themselves to the gradient of the loss function. Specifically, it applies gradient clipping to the gradient of the regularized loss function, allowing the model parameters to adapt freely to the new task while still preventing over-fitting through controlled gradient updates. This resolves the contradiction by maintaining reliability without damaging adaptability.
2Adaptability or versatility
If standard fine-tuning is applied to target task with limited samples, then model adapts to new task, but over-fitting occurs and generalization capability is lost
Solution Approach 1:
The patent introduces the regularized loss function as an intermediary between the standard cross-entropy loss and the parameter updates. The regularized loss function combines the cross-entropy loss with an additional regularization term, and gradient clipping is applied to this regularized loss. This intermediary mechanism allows the model to adapt to the target task while preventing over-fitting, thus maintaining generalization capability.
3Reliability
If SPAR method is used to prevent over-fitting by constraining parameters near source model, then generalization is improved, but the model cannot learn high-quality representations for target task
Solution Approach 1:
The patent makes the constraint mechanism dynamic by applying gradient clipping adaptively during training. Unlike static parameter constraints in SPAR that prevent the model from moving away from source parameters, the dynamic gradient clipping allows parameters to evolve freely as long as the gradient updates remain within acceptable bounds. This enables the model to learn high-quality representations for the target task while maintaining generalization capability.
Data Source
AI summary
A method, apparatus, device and readable medium for transfer learning in machine learning are provided. The method includes: constructing a target model according to the number of classes to be achieved by a target task and a duly-trained source model; obtaining a value of a regularized loss function of the corresponding target model and a value of a cross-entropy loss function of the target model, based on sets of training data in a training dataset of the target task; according to the value of the regularized loss function and the value of the cross-entropy loss function corresponding to each set of training data, updating parameters in the target model by a gradient descent method to implement the training of the target model. The above technical solution avoids excessive constraints on parameters in the prior art, thereby refraining from damaging the training effect of the source model on the target task.