Learning Device for Interrelated Feature Representation Across Data Forms
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Feature value vectors derived by common autoencoders do not effectively interrelate different forms of the same target, leading to poor identification performance when the target is viewed from varying directions or angles, as the vectors are unrelated even when the data represents the same target.
Innovation Solution
A learning device that acquires and processes data with different forms of a target, using an encoder to derive feature values, a conversion unit to transform these values, and a decoder to generate new data, with parameter updating based on comparisons between the original and transformed data to ensure interrelated feature representation across different forms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a common autoencoder is used to derive feature value vectors, then the encoding process is simple and straightforward, but the feature values for the same target in different forms become unrelated, leading to poor identification performance
Solution Approach 1:
The patent introduces a conversion unit as an intermediary component between the encoder and decoder. This conversion unit transforms the first feature value (from first data) into a second feature value (for second data showing different forms of the same target), enabling the system to maintain interrelated feature representations across different target forms while preserving the simplicity of the autoencoder structure
Solution Approach 2:
The patent segments the traditional autoencoder structure by inserting a conversion unit between the encoder and decoder. This segmentation allows the system to handle different target forms separately through the conversion unit while maintaining the overall autoencoder framework, thus improving identification performance without completely redesigning the encoding process
2Reliability
If training data completely covering various forms of a target is prepared, then identification accuracy for all forms can be improved, but data preparation becomes difficult and time-consuming
Solution Approach 1:
The patent implements preliminary action by pre-training the autoencoder with a subset of training data covering various target forms. The conversion unit is trained to map feature values across different forms, enabling the system to handle unseen forms without requiring exhaustive training data preparation. This preliminary setup allows the system to generalize to new forms efficiently
Solution Approach 2:
The patent utilizes parameter changes by training the conversion unit with transformation parameters that represent different target forms (such as rotation angles, scaling factors, or other form-differentiating parameters). By learning these parameter transformations, the system can generalize to new forms without needing explicit training examples for each possible form, significantly reducing data preparation requirements
Data Source
AI summary
Provided is a learning device that can generate a feature deriving device capable of deriving, for an identical object, feature amounts which respectively express a feature of the object in different forms and which are mutually related. This learning device comprises: an acquisition unit that acquires first data and second data, with different forms of the object recorded therein; an encoder that derives a first feature amount from the first data; a conversion unit that converts the first feature amount to a second feature amount; a decoder that generates third data from the second feature amount; and a parameter updating unit that updates, on the basis of a comparison between the second data and the third data, the value of a parameter used in the derivation of the first feature amount, and the value of a parameter used in the generation of the third data.