AI Model Adaptation via Dynamic Determination Modes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning models, particularly neural networks, face challenges in adapting to new data and are biased towards best-performing classes, leading to accuracy issues and reliance on large labeled datasets, with active learning methods requiring frequent model resets and batch-based data classification, which is resource-intensive and inefficient.
Innovation Solution
An electronic data processing device with an acquisition module, estimation module, determination module, and learning module that selects between assisted, semi-supervised, and self-supervised determination modes based on prediction uncertainty, allowing for continuous local learning and synaptic weight modification using backpropagation, enabling more reliable task determination and adaptation to new data without requiring extensive labeled data or remote server connections.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If active learning methods are used to reduce dependence on labeled data, then the need for large labeled datasets is reduced, but the model requires frequent resets and batch-based processing which increases resource consumption and processing time
Solution Approach 1:
The patent implements dynamic determination modes that adapt between assisted, semi-supervised, and self-supervised approaches based on prediction uncertainty. This allows the system to process data continuously without batch-based resets, improving productivity while reducing labeled data requirements through intelligent mode switching
Solution Approach 2:
The patent enables continuous learning by processing data streams without interruption or model resets. The determination module continuously adapts the prediction model using available data in appropriate modes, eliminating the batch-based processing interruptions inherent in traditional active learning methods
2Adaptability or versatility
If the prediction model continuously adapts to new data through local learning, then the model remains updated and relevant, but more computational resources are consumed for synaptic weight modifications
Solution Approach 1:
The patent implements local learning by modifying only the synaptic weights of the prediction model locally rather than retraining the entire model. The learning module performs targeted weight adjustments based on determination modes and prediction uncertainty, reducing overall computational resource consumption while maintaining adaptability
Solution Approach 2:
The patent changes the parameters being optimized from complete model retraining to specific synaptic weight modifications. By adjusting only the necessary weights based on determination modes and uncertainty levels, the system achieves continuous adaptation with reduced computational overhead
3Reliability
If multiple determination modes are implemented to handle different data scenarios, then the system becomes more flexible and accurate, but the device complexity increases
Solution Approach 1:
The patent uses dynamic mode selection where the determination module automatically chooses between assisted, semi-supervised, and self-supervised modes based on prediction uncertainty thresholds. This dynamic approach provides multiple determination strategies without requiring complex manual configuration or switching mechanisms
Solution Approach 2:
The patent implements feedback loops where prediction uncertainty results from the estimation module feed back to the determination module, which then selects appropriate determination modes. This feedback mechanism coordinates the multiple modes systematically, managing complexity through automated decision-making based on performance metrics
Data Source
Figure 1
Figure 2
Figure 3
AI summary
This processing device is connected to a sensor (15) and includes a data acquisition module (22), a task estimation module (24) for estimating a task associated with the data via a prediction model (60), and a task determination module (28) for determining a resulting task for the data. The determination module calculates a quantity (E) representative of the estimate; then selects—based on the quantity—one of two assisted (ACL) and supervised (SSL, AUL) determination modes; and determines—according to this mode—the resulting task. In assisted determination mode, the determination module sends a request to a user for a task for the data. The device includes a learning module (30) for modifying synaptic weights of the prediction model via backpropagation to calculate the gradient of each weight according to a cost function dependent on the quantity.