Learning Device Variable Derivation for Classification Precision

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing classifiers optimized by common autoencoders struggle to interrelate different forms of the same target, leading to poor identification performance when faced with variations in direction or angle, even with small datasets.

Innovation Solution

A learning device that performs variable derivation learning and classification learning, using a variable derivation unit to generate interrelated latent variable vectors representing different forms of a target through variable conversion, and a classification learning unit to ensure correct classification across various forms.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a common autoencoder is used to derive feature value vectors, then the classification precision for a specific direction is improved, but the ability to identify targets in different directions or angles deteriorates

Engineering Contradiction:
Improveclassification precisionVSAvoididentification capability across different forms
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent applies dynamics by making the latent variable vector representation adaptive to different forms of the same target. The encoder dynamically adjusts the encoding process based on the input data's characteristics, allowing the system to maintain high classification precision across various directions and angles rather than being fixed to a single viewpoint.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent changes the parameter representation in the latent variable vector space to accommodate different forms of targets. By transforming and rerepresenting the latent variables according to the specific characteristics of each target form, the system achieves both high precision for individual classes and versatility across different representations of the same target.

Inventive Principle:
Principle #35Parameter changes

2Reliability

If training data completely covering various forms is used, then the identification performance on various forms is improved, but the difficulty of data preparation and the quantity of data required worsen

Engineering Contradiction:
Improveidentification performanceVSAvoiddata preparation complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent uses copying by generating synthetic training data that replicates the characteristics of real target data. The encoder learns to create latent variable vectors that capture the essential features of targets in various forms, allowing the system to train on a smaller, more manageable dataset while still achieving high identification performance across different forms.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent applies preliminary action by pre-processing and transforming the training data into a standardized latent variable representation before classification. This preliminary encoding process simplifies the data preparation requirements by automatically extracting and normalizing features, reducing the manual effort needed to prepare comprehensive training datasets.

Inventive Principle:
Principle #10Preliminary action

3Ease of manufacture

If the number of training data samples is reduced, then the ease of data collection is improved, but the classification accuracy on unseen forms deteriorates

Engineering Contradiction:
Improveease of data collectionVSAvoidclassification accuracy
Core Design Contradiction:
Ease of manufactureVSMeasurement precision

Solution Approach 1:

The patent generates synthetic training samples through the encoder that copy the essential characteristics of real target data. This allows the system to achieve high classification accuracy with fewer real training samples, as the synthetic data complements and expands the training set without requiring additional data collection effort.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The encoder performs self-service by automatically extracting and representing the essential features of training data into latent variable vectors. This self-encoding capability eliminates the need for manual feature engineering and reduces the dependency on large quantities of manually collected and processed training data, improving both ease of data collection and classification accuracy.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP3726463B1Learning device, learning method, sorting method, and storage medium
Publication Date: 2025.02.19 NEC CORP
  • EP3726463B1 patent drawingFigure 1A~1B
  • EP3726463B1 patent drawingFigure 1C
  • EP3726463B1 patent drawingFigure 2

AI summary

Provided is a learning device that can generate a discriminator which makes it possible to identify an object having various forms, even if there are few samples of data with the object recorded therein. A learning device according to one embodiment comprises: an acquisition unit that acquires a first feature amount derived by an encoder from data with an identification object recorded therein, the encoder being configured so as to derive, from data with the identical object in various forms recorded therein, feature amounts which are mutually convertible by a conversion using a conversion parameter that takes a value according to the difference in the forms; a conversion unit that generates a second feature amount by performing a conversion on the first feature amount using the conversion parameter value; and a parameter updating unit that updates the value of a sorting parameter used in sorting by a sorting means, which is configured to sort second feature amounts as input, such that if the second feature amount has been input, the sorting means outputs a result indicating, as a sorting destination, a class associated with the identification object.