Machine Learning Training With Learning-Parameter Feedback for Rare Classes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning training methods for technical systems face inefficiencies when dealing with rare recognition classes, where training data is scarce and recognition reliability is difficult to improve, often requiring extensive data collection before recognizing these classes reliably.
Innovation Solution
A method that determines the value distribution of learning parameters during training, using a Bayesian neural network, to generate a continuation signal for deciding whether to continue with the same or switch to different training data, optimizing the training process by requesting qualitatively different data early on.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If training data is continuously increased to improve recognition reliability for rare classes, then recognition reliability may improve, but training time and training effort increase considerably
Solution Approach 1:
The patent applies preliminary action by continuously monitoring the value distribution of learning parameters during training to predict future recognition reliability before actually collecting more training data. This allows the system to determine in advance whether additional training data will be beneficial, avoiding unnecessary data collection and training time expenditure.
Solution Approach 2:
The patent implements feedback by using the value distribution of learning parameters as a real-time indicator to guide training decisions. The monitoring mechanism provides continuous feedback about training effectiveness, enabling dynamic adjustment of training strategies based on actual learning progress rather than fixed predetermined schedules.
2Reliability
If training is continued with different sensor data when current data is insufficient, then recognition reliability may improve, but training effort increases considerably
Solution Approach 1:
The patent applies preliminary action by analyzing the value distribution of learning parameters to predict whether training will be effective before actually proceeding with additional training data collection. This preliminary assessment prevents wasted effort on training that is unlikely to improve recognition reliability.
Solution Approach 2:
The patent employs parameter changes by using the value distribution characteristics of learning parameters as a dynamic criterion for deciding whether to continue or stop training. This parameter-based decision mechanism allows the system to adapt training strategies based on actual learning state rather than fixed data quantity thresholds.
Data Source
AI summary
To train a machine learning routine (BNN), a sequence of first training data (PIC) is read in through the machine learning routine. The machine learning routine is trained using the first training data, wherein a plurality of learning parameters (LP) of the machine learning routine is set by the training. Furthermore, a value distribution (VLP) of the learning parameters, which occurs during the training, is determined and a continuation signal (CN) is generated on the basis of the determined value distribution of the learning parameters. Depending on the continuation signal, the training is then continued with a further sequence of the first training data or other training data (PIC2) are requested for the training.