On-Device User Recognition via Dual Neural Network Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing user recognition systems face challenges in accurately identifying users with similar features, such as family members, due to the mismatch between training data and actual user data, leading to poor recognition performance.
Innovation Solution
The proposed solution involves a dual neural network approach, where a first neural network is pre-trained on a large user database with set parameters and a second neural network is trained on-device using user data and reference data from generalized users, allowing for on-device training to adjust parameters and improve recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If a single neural network is trained on a large user database, then the system can process many users, but recognition accuracy deteriorates for users with similar features due to data mismatch
Solution Approach 1:
The patent divides the neural network into two separate networks: a first neural network trained on a large user database for general feature extraction, and a second neural network trained on-device on specific user data for accurate recognition. This segmentation allows each network to specialize - the first handles diversity while the second handles precision for individual users, resolving the contradiction between handling many users and maintaining accuracy.
Solution Approach 2:
The first neural network acts as an intermediary that processes raw user data and extracts general features, which are then passed to the second neural network for refined processing. This intermediary structure allows the system to benefit from both large-scale training data and user-specific customization, maintaining accuracy while handling diverse user populations.
2Productivity
If parameters are fixed after pretraining, then computational efficiency improves, but adaptability to individual users deteriorates
Solution Approach 1:
The first neural network performs preliminary training on a large user database before deployment, extracting generalizable features and establishing a solid foundation. This preliminary action allows the second network to focus only on user-specific adaptations during on-device training, improving both efficiency and adaptability by separating general learning from specific customization.
Solution Approach 2:
The patent implements a dynamic parameter structure where the first neural network has fixed parameters for stability and efficiency, while the second neural network has adjustable parameters that adapt to individual users during on-device training. This dynamic approach allows the system to balance computational efficiency with user-specific adaptability.
3Measurement precision
If on-device training is performed with limited user data, then user-specific accuracy improves, but training reliability deteriorates due to data scarcity
Solution Approach 1:
The first neural network serves as an intermediary that pre-processes and enriches limited user data by transferring general knowledge from the large training database. This intermediary processing allows the second network to perform reliable on-device training even with limited user data, as the first network has already extracted robust general features that complement the scarce user-specific data.
Solution Approach 2:
The patent changes the training parameters and data composition for the two networks: the first network is trained on large-scale diverse data with specific loss functions, while the second network is trained on smaller user-specific data with different optimization parameters. This parameter differentiation allows each network to be optimally trained for its purpose, improving reliability despite data limitations.
Data Source
AI summary
An on-device training-based user recognition method includes performing on-device training on a feature extractor based on reference data corresponding to generalized users and user data, determining a registration feature vector based on an output from the feature extractor in response to the input of the user data, determining a test feature vector based on an output from the feature extractor in response to an input of test data, and performing user recognition on a test user based on a result of comparing the registration feature vector to the test feature vector.