Classifier Segmentation for Unknown Class Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional classifiers face challenges in detecting unknown classes and initializing classifiers for samples that do not belong to trained classes, leading to difficulties in accurately categorizing data and high computational complexity.

Innovation Solution

A method involving the generation of a first classifier with a dimension of at least two and a second classifier to determine whether input data belongs to known or unknown classes, using synthetic negative data and binary classifiers to differentiate between known and unknown classes.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If conventional classifiers are used for classification, then classification capability is provided, but computational complexity becomes prohibitively high and unknown classes cannot be detected

Engineering Contradiction:
Improveclassification accuracyVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The classification system is segmented into two distinct stages: a first classifier for known classes and a second classifier for unknown class detection. This segmentation allows each classifier to be optimized for its specific function, reducing overall computational complexity while maintaining reliability.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

A second classifier acts as an intermediary between the first classifier and the final classification output. This intermediary detects unknown classes by analyzing the confidence scores or probability distributions from the first classifier, enabling reliable detection without requiring the first classifier to handle all possible classes.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Ease of manufacture

If conventional classifiers are trained only on known classes, then training is simplified, but detection of unknown classes becomes impossible

Engineering Contradiction:
Improvetraining simplicityVSAvoidunknown class detection capability
Core Design Contradiction:
Ease of manufactureVSAdaptability or versatility

Solution Approach 1:

The system performs preliminary classification with the first classifier trained only on known classes, then uses the second classifier to preliminarily detect potential unknown classes before final classification. This preliminary action maintains training simplicity while enabling adaptability to unknown classes.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The second classifier provides feedback about unknown class detection to the overall system, allowing the first classifier to maintain its simplified training on known classes while the feedback mechanism enables detection of unknown classes through the probability distribution analysis.

Inventive Principle:
Principle #23Feedback

3Adaptability or versatility

If deep learning architectures are used, then flexibility and fine-tuning capability are improved, but training becomes challenging and computational requirements increase

Engineering Contradiction:
Improvefine-tuning capabilityVSAvoidtraining difficulty
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The deep learning architecture is segmented into two separate classifier components, each with distinct training objectives. The first classifier is trained on known classes with standard supervised learning, while the second classifier is trained to detect unknown classes, dividing the training complexity into manageable parts.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Instead of training a single comprehensive classifier to handle all possible classes including unknown ones, the system uses partial action by training the first classifier only on known classes and using the second classifier to handle the unknown class detection, avoiding the excessive training complexity of a monolithic approach.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3332362B1Detection of unknown classes and initialization of classifiers for unknown classes
Publication Date: 2024.10.30 QUALCOMM INC
  • EP3332362B1 patent drawingFigure 1
  • EP3332362B1 patent drawingFigure 2
  • EP3332362B1 patent drawingFigure 3A

AI summary

A method of detecting unknown classes is presented and includes generating a first classifier for multiple first classes. In one configuration, an output of the first classifier has a dimension of at least two. The method also includes designing a second classifier to receive the output of the first classifier to decide whether input data belongs to the multiple first classes or at least one second class.