Test-Time Adaptation for Image Classifiers Using Non-Saturating Loss

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Deep neural networks often perform poorly on data distributions different from their training data, leading to poor robustness and adaptability in real-world deployments where data shifts are common and unpredictable.

Innovation Solution

A method for fully test-time adaptation using a novel loss function that replaces entropy with a non-saturating surrogate and adds a diversity regularizer through batch-wise entropy maximization, along with an input transformation module that can undo domain shifts, allowing adaptation without labeled target data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If entropy minimization is used for test-time adaptation, then performance on shifted distributions improves, but the method suffers from premature convergence and instability

Engineering Contradiction:
Improveperformance on shifted distributionsVSAvoidconvergence stability
Core Design Contradiction:
ReliabilityVSStability of the object's composition

Solution Approach 1:

The patent replaces the entropy loss function with a non-saturating surrogate loss function that maintains gradient flow throughout training. This parameter change in the loss function prevents premature convergence by ensuring that gradients remain informative even when prediction confidence is high, thereby stabilizing the adaptation process while maintaining improvement on shifted distributions.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent introduces a diversity regularizer based on batch-wise entropy maximization as an intermediary component that works alongside the non-saturating loss. This regularizer acts as a mediator to prevent collapse to trivial solutions by encouraging diverse predictions across batches, thus stabilizing convergence without sacrificing performance on shifted distributions.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If labeled target data is collected for adaptation, then model robustness to domain shifts improves, but the process becomes time-consuming and expensive

Engineering Contradiction:
Improverobustness to domain shiftsVSAvoiddata collection and labeling time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent enables the model to adapt to domain shifts using only unlabeled target data through self-supervised learning. The non-saturating loss function and diversity regularizer work together to provide self-supervision signals from the unlabeled data itself, eliminating the need for time-consuming manual labeling while maintaining robustness improvement.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent performs preliminary adaptation during the test time itself using the non-saturating loss function on unlabeled target data. This preliminary action of adapting to the target distribution before actual deployment allows the model to become robust to domain shifts without requiring prior labeled target data collection and annotation.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If full retraining is performed to adapt to new distributions, then classification accuracy is maintained, but computational resources and time are significantly consumed

Engineering Contradiction:
Improveclassification accuracyVSAvoidadaptation speed
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent applies partial action by using the non-saturating loss function with only unlabeled target data for adaptation, which requires significantly fewer computational resources than full retraining. This partial approach maintains sufficient classification accuracy by focusing adaptation efforts on the critical aspect of domain shift robustness rather than requiring complete retraining on labeled data.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The patent changes the loss function parameter from traditional saturating entropy to a non-saturating surrogate, which enables efficient adaptation with unlabeled data only. This parameter change allows the model to achieve good adaptation results without the computational burden of full retraining, thereby improving productivity while maintaining acceptable accuracy.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4105847A1Device and method to adapt a pretrained machine learning system to target data that has different distribution than the training data without the necessity of human annotations on target data
Publication Date: 2022.12.21 ROBERT BOSCH GMBH
  • EP4105847A1 patent drawingFigure 1
  • EP4105847A1 patent drawingFigure 2A~2B
  • EP4105847A1 patent drawingFigure 3

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. Furthermore, an input transformation module (2) for partly undo the distribution shift between the first and second training data set is proposed.