Dummy Node Framework for Neural Network Continual Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing user authentication techniques, such as password-based systems and two-factor authentication, face limitations including user experience issues, password complexity, and inefficiencies in training machine-learning models with additional sensor data, leading to time-consuming retraining processes, especially on mobile devices with limited resources.
Innovation Solution
A system that uses a multilayer neural network model with dummy nodes to update weights efficiently, allowing for authentication of both initial and new users using sensor data, employing techniques like least-squares regression and federated learning to maintain confidentiality and avoid catastrophic forgetting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the system uses traditional retraining methods to incorporate additional sensor data from new users, then the model can recognize new user classes, but the training process becomes time-consuming and computationally intensive
Solution Approach 1:
The patent segments the neural network into multiple layers, where only the last layer is retrained when new user data becomes available. This segmentation allows the system to update the model for new users without retraining the entire network, significantly reducing training time while maintaining the ability to recognize both old and new user classes.
Solution Approach 2:
The patent introduces dummy nodes in advance during the initial model training phase, before actual new user data is available. These dummy nodes are pre-positioned in the last layer and will later be associated with new users when their data becomes available, eliminating the need for structural modifications and enabling rapid adaptation.
2Reliability
If the system retrains the entire neural network model from scratch to accommodate new sensor data, then the model accuracy is maintained, but the computational resources and time required increase significantly
Solution Approach 1:
The patent divides the training process into segments where only the last layer of the neural network is retrained when new user data arrives. This segmentation maintains model accuracy for new users while preserving the already-trained weights in earlier layers, dramatically improving training efficiency and reducing computational resource requirements.
Solution Approach 2:
The patent discards the practice of retraining entire models from scratch. Instead, it recycles and reuses the previously trained weights in the earlier layers of the neural network, only updating the necessary last layer with new user data, thereby maintaining accuracy while boosting productivity.
3Ease of operation
If the system uses dummy nodes in the neural network model, then the model structure remains fixed and updates are faster, but additional computational overhead is introduced during the updating process
Solution Approach 1:
The patent uses dummy nodes that are essentially placeholder copies of potential user class representations. These dummy nodes are created in advance and simply mapped to new users when data becomes available, adding minimal structural complexity while enabling easy model updates without requiring structural modifications.
Data Source
AI summary
The disclosed embodiments provide a system that identifies a user of an electronic device. During a training mode, the system uses an initial training data set, comprising sensor data from electronic devices associated with a set of initial users, to train a multilayer neural network model to authenticate the initial users. Next, the system uses an additional training data set, which includes sensor data from electronic devices associated with one or more new users, to update a portion of the weights in the trained model so that the updated model can be used to authenticate both the initial users and the one or more new users. During a subsequent mode, the system uses the updated model to authenticate a user of the electronic device based on sensor data contemporaneously received from the electronic device.