Classifier Training with Selected Negative Classes

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Classifiers trained on specific data can only recognize known classes, leading to unpredictable results when encountering unknown images, as they lack the ability to distinguish between known and unknown classes effectively.

Innovation Solution

The method involves selecting negative classes from a broader candidate set that are similar but not too similar or too different from the target classes, merging similar classes, and training the classifier on a combined set of these negative and target classes to enhance its recognition capabilities and robustness.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If a classifier is trained only on provided target class data, then the training process is simple and fast, but the classifier cannot recognize unknown classes and produces unpredictable results

Engineering Contradiction:

Improveclassifier accuracy on unknown classesVSAvoidtraining data complexity

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The system performs preliminary action by automatically selecting and adding negative classes from candidate background classes before the actual classifier training. This preprocessing step prepares the training data by identifying suitable negative classes that are similar to target classes but not too similar, ensuring the classifier will be exposed to relevant unknown classes during training without manual intervention.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system implements self-service by automatically selecting negative classes from candidate background classes based on similarity metrics, without requiring manual curation. The automated selection process evaluates candidate classes and identifies those that are similar to target classes but not too similar, allowing the system to prepare its own training data independently.

Inventive Principle:

Principle #25Self-service

2Reliability

If a classifier is trained on a wide variety of classes, then the classifier becomes more robust to unknown inputs, but the training time and computational resources increase

Engineering Contradiction:

Improveclassifier robustnessVSAvoidtraining time

Core Design Contradiction:

ReliabilityVSLoss of time

Solution Approach 1:

The system applies local quality by selecting negative classes with specific local characteristics - they are similar to target classes but not too similar. This targeted selection focuses training on locally relevant unknown classes rather than uniformly sampling from all possible classes, making training more efficient while maintaining robustness for the specific application domain.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The system changes parameters by adjusting the similarity threshold and selection criteria for negative classes. By modifying these parameters, the system can control the diversity and relevance of negative classes in the training set, balancing between robustness and training efficiency based on specific requirements.

Inventive Principle:

Principle #35Parameter changes

3Adaptability or versatility

If random background classes are selected for training, then diverse unknown classes are covered, but classes that are too different or too similar to target classes are included reducing training effectiveness

Engineering Contradiction:

Improvecoverage of unknown classesVSAvoidtraining data quality

Core Design Contradiction:

Adaptability or versatilityVSManufacturing precision

Solution Approach 1:

The system uses feedback by evaluating candidate background classes against similarity metrics relative to target classes. Classes that are too similar or too different are filtered out based on this feedback, ensuring only appropriately similar negative classes are selected. This feedback mechanism maintains high training data quality while achieving good coverage of relevant unknown classes.

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The system changes parameters by adjusting similarity thresholds and selection criteria to optimize the balance between coverage and quality. By modifying these parameters, the system can control which candidate classes are included as negative classes, ensuring they provide diverse coverage without including classes that are too similar or too different from target classes.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS11829848B2Adding negative classes for training classifier

Publication Date: 2023.11.28 MICROSOFT TECHNOLOGY LICENSING LLC

AI summary

A method includes obtaining training data for a classifier, the training data comprises one or more target classes, obtaining candidate background classes, selecting negative classes from the candidate background classes, wherein the negative classes exclude candidate background classes that are close to the target classes, wherein the negative classes exclude candidate background classes that are very different from the target classes, and wherein the negative classes include candidate background classes that are similar to the target classes, and training the classifier on a combined set of the selected negative classes and target classes.