Pretrained ML Adaptation via Non-Saturating Loss and Input Transformation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks often perform poorly on data distributions that differ from the training data distribution, leading to challenges in adapting to domain shifts without labeled target data.
Innovation Solution
A novel loss function is introduced that replaces entropy with a non-saturating surrogate and adds a diversity regularizer based on batch-wise entropy maximization to prevent premature convergence. Additionally, an input transformation module is prepended to the network to partially undo test-time distribution shifts, which can be learned end-to-end without target domain labels or source domain data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If entropy minimization is used for test-time adaptation, then performance on shifted distributions is improved, but premature convergence and instability occur
Solution Approach 1:
The patent replaces the entropy loss function with a non-saturating surrogate loss function that maintains gradient flow during convergence. This parameter change in the objective function prevents premature convergence while maintaining the ability to adapt to distribution shifts, directly resolving the contradiction between improvement and stability.
Solution Approach 2:
The patent introduces a diversity regularizer as an intermediary component that works alongside the loss function. This regularizer prevents collapse to trivial solutions by maintaining diversity in predictions, thereby stabilizing the adaptation process while preserving performance improvement on shifted distributions.
2Measurement precision
If labeled target data is used for adaptation, then classification accuracy is improved, but labeling cost and time increase
Solution Approach 1:
The patent enables the model to adapt to target distributions using only unlabeled target data through self-supervised learning. The model serves itself by learning from the structure and statistics of unlabeled data without requiring external labeling, thereby maintaining classification accuracy while eliminating labeling time and cost.
3Measurement precision
If full retraining is performed to restore accuracy on target distribution, then classification accuracy is improved, but computational cost and time increase
Solution Approach 1:
The patent performs only partial adaptation by updating selected parameters of the pretrained model using a small number of epochs on unlabeled target data. This partial action is sufficient to restore accuracy on the target distribution without requiring full retraining, thereby maintaining productivity while achieving the desired accuracy improvement.
Data Source
AI summary
A computer-implemented method for adapting a pretrained machine learning system, which has been trained on a first training data set, to a second dataset, wherein the second dataset has different characteristics than the first data set. An input transformation module for partly undoing the distribution shift between the first and second training data set is provided.