Neural Network Weight Update via Confidence-Based Data Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing Deep Neural Network (DNN) image recognition systems face challenges in achieving sufficient recognition accuracy, especially when using data with unknown correct answers, and require extensive learning data and time-consuming learning processes.
Innovation Solution
A learning utilization system comprising a utilizing device and a learning device, where the utilizing device acquires target data, performs inference using a neural network, determines the certainty of the inference result, and transmits data with low certainty to the learning device for relearning. The learning device updates its neural network weights using the transmitted data and sends the updated coefficients back to the utilizing device for application.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large amount of learning data is used to achieve sufficient accuracy in DNN, then recognition accuracy is improved, but the learning process becomes substantially time-consuming and requires large memory size
Solution Approach 1:
The patent extracts only the necessary learning data by selecting data where the inference network's confidence score is below a predetermined threshold. This extraction approach filters out data that would not contribute to improving recognition accuracy in the specific installation environment, thereby reducing the amount of data needed for learning while maintaining accuracy improvements.
Solution Approach 2:
The patent applies local quality by adapting the learning process to the specific installation environment. Instead of using all collected data uniformly, the system selectively uses data based on the confidence scores generated by the inference network in that particular environment, making the learning process locally optimized for each installation scenario.
2Adaptability or versatility
If a DNN model learned by widely collected data is applied to a particular installation environment, then general data coverage is improved, but recognition accuracy in the specific environment deteriorates
Solution Approach 1:
The patent implements dynamics by making the learning process adaptive to each installation environment. The system dynamically selects learning data based on the confidence scores generated by the inference network in the specific environment, allowing the model to adapt to local characteristics while maintaining the benefit of widely collected data.
Solution Approach 2:
The patent applies local quality by tailoring the learning process to each installation environment's specific needs. The system uses confidence scores from the inference network to identify which data points are most useful for improving accuracy in that particular environment, thereby localizing the learning process to match environmental-specific requirements.
3Productivity
If learning data with unknown correct answers is used, then data utilization efficiency is improved, but the reliability of learning results deteriorates
Solution Approach 1:
The patent applies feedback by using the inference network's confidence scores as a feedback mechanism to select learning data. The confidence score indicates the network's uncertainty about a data point, and this feedback is used to selectively include uncertain data in the learning process, thereby improving data utilization efficiency while maintaining learning reliability through structured selection.
Solution Approach 2:
The patent implements preliminary action by performing inference first to generate confidence scores before selecting data for learning. This preliminary inference step allows the system to pre-filter and select data points that are most suitable for learning, ensuring that data with unknown correct answers is selectively used only when beneficial, thus improving both efficiency and reliability.
Data Source
AI summary
A utilizing device includes a utilizing-side inference unit that uses a utilizing-side inference network to perform inference from target data, and a utilizing-side transmitting unit that transmits the target data to a learning device when the degree of certainty of the inference result inferred by the utilizing-side inference network is lower than a predetermined criterion. The learning device includes a learning-side inference unit that uses a learning network functioning as a teacher model of a utilizing-side inference network to perform inference from the received target data to generate learning data, a learning unit that generates updated weighting coefficient information indicating weighting coefficients updated by relearning the learning-side inference network which has the same network structure as the utilizing-side inference network by using the learning data, and a learning-side transmitting unit that transmits the updated weighting coefficient information to the utilizing device to update the utilizing-side inference network.