Neural Network Training With Secondary Phase For Misestimated Values
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for training artificial neural networks to identify digital uses are inadequate in terms of reliability and accuracy, as they often rely on personal data and non-immutable information, which compromises user privacy and lack robustness in identifying multiple properties describing digital behavior and environment.
Innovation Solution
A method involving primary and secondary training phases, where the neural network is trained using gradient backpropagation to identify target property values, with the secondary training phase focusing on increasing learning rates for neurons associated with misestimated values, thereby enhancing the model's robustness and accuracy in identifying digital uses without relying on personal data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a neural network is trained using conventional methods on available data, then the training process is simple and fast, but the reliability and accuracy of identifying digital uses is insufficient
Solution Approach 1:
The training process is divided into two distinct phases: primary training using initial training data, and secondary training using specifically selected confusion data. This segmentation allows the system to first establish basic identification capabilities and then refine them by addressing specific weaknesses, thereby improving reliability without overwhelming complexity
Solution Approach 2:
Before the secondary training phase, the system performs preliminary analysis to identify confusion values (incorrect predictions) from the primary training results. This preliminary action enables targeted selection of training data that specifically addresses identified weaknesses, making the subsequent training more efficient and effective
2Measurement precision
If the neural network uses uniform learning rates for all neurons, then the training is straightforward, but the accuracy for misestimated values cannot be improved
Solution Approach 1:
The system applies different learning rates to different neurons in the output layer based on their specific performance characteristics. Neurons associated with confusion values (incorrect predictions) receive higher learning rates, while other neurons maintain standard learning rates. This localized differentiation improves prediction accuracy for problematic cases without unnecessarily complicating the overall training management
3Quantity of substance
If personal data is used for training the neural network, then the training data is readily available, but user privacy is compromised
Solution Approach 1:
The system extracts and removes personally identifiable information from the training data while retaining the behavioral patterns and features necessary for digital use identification. This extraction process allows the system to maintain adequate training data quantity for effective model training while eliminating privacy-harmful elements
Data Source
AI summary
A method for training an artificial neural network so that the artificial neural network identifies a property value among a plurality of property values, each property being able to take at least two different values. The method includes: a primary training including training the neural network to identify at least one target value; and a secondary training for detecting weaknesses of the model trained during the primary training and reinforcing this model by increasing the learning rate of output neurons of the network that are associated with property values most often misestimated.