Electronic Device Biometric Learning for Label-Efficient Latent Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing prediction models face challenges in accurately predicting latent physiological variables for users due to the difficulty in collecting and optimizing labeling data, leading to inefficient user-specific predictions.
Innovation Solution
An electronic device utilizes an unsupervised learning model to predict measured variables and a supervised learning model to predict latent variables, operating independently of labeling data, and uses feedback from supervised learning to optimize predictions for individual users.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a prediction model uses labeled dataset for learning, then prediction accuracy can be improved, but data collection and labeling complexity increases significantly
Solution Approach 1:
The patent divides the prediction model into two separate models: a supervised learning model that uses labeled datasets to predict latent variables, and an unsupervised learning model that processes biometric signals without requiring labeling. This segmentation allows the system to leverage labeled data where it provides maximum value (latent variable prediction) while avoiding the complexity of labeling routine measurements.
Solution Approach 2:
The system performs preliminary unsupervised learning to predict measured variables from biometric signals before applying supervised learning to predict latent variables. This preliminary action reduces the burden of data collection and labeling by pre-processing the biometric data and identifying patterns that can be used as features for the supervised model.
2Device complexity
If a single prediction model is used for both measured variables and latent variables, then device complexity is reduced, but prediction accuracy for user-specific characteristics deteriorates
Solution Approach 1:
The patent implements segmentation by creating distinct supervised and unsupervised learning models for different prediction tasks. The unsupervised model handles measured variable prediction from raw biometric signals, while the supervised model focuses on latent variable prediction using labeled datasets. This division enables each model to be optimized for its specific function, improving overall user-specific prediction accuracy.
Solution Approach 2:
The system adds a dimensional layer by introducing both supervised and unsupervised learning components working in parallel. The unsupervised model operates on raw biometric signal dimensions, while the supervised model operates on processed feature dimensions with labeled outcomes, creating a multi-dimensional prediction architecture that captures different aspects of user characteristics.
3Adaptability or versatility
If labeled datasets are collected from multiple users, then model generalization improves, but data privacy and security requirements increase
Solution Approach 1:
The unsupervised learning model serves as an intermediary that processes raw biometric signals and extracts features without requiring personal identifiable information or detailed labeling. This intermediary layer enables the system to learn from multiple users' data patterns while minimizing the exposure of sensitive personal information, as the supervised model only needs anonymized feature representations rather than raw biometric data.
Data Source
AI summary
An electronic device includes a sensor configured to acquire a biometric signal of a user; a memory configured to store a measured variable of the user; and a processor configured to: receive biometric data of a plurality of other users and labeling data corresponding to the biometric data of the plurality of other users; predict, using an unsupervised learning model, biometric data of the user based on the biometric signal of the user; predict, using a supervised learning model, a biometric condition of the user based on the predicted biometric data of the user, the biometric data of the plurality of other users, and the labeling data.