Automated Neural Network Retraining via Failed Data Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks require continuous retraining with updated data to maintain optimal performance, but this process is often manual and time-consuming, limiting their application in real-world scenarios where speed and iterative updates are necessary.
Innovation Solution
An automated method for selecting and annotating input data that fails processing by the neural network, enriching the training sample data set, and iteratively retraining the network without human intervention, using a control unit, selecting, annotating, and training units within a computer server to enhance network accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual retraining process is used to improve neural network performance, then network accuracy can be improved, but the process becomes time-consuming and requires human involvement
Solution Approach 1:
The system automatically selects failed input data, obtains annotated data, and retrains the neural network without human intervention. The control unit triggers the selecting unit to identify failed processing cases, the annotating unit to obtain corrected data, and the training unit to perform iterative retraining, enabling the system to self-improve its accuracy automatically.
Solution Approach 2:
The system uses the processing results of the neural network as feedback to identify which input data failed processing. This feedback loop allows the system to continuously select problematic cases, obtain their correct annotations, and retrain on these specific examples, creating a closed-loop improvement mechanism that targets actual performance weaknesses.
2Adaptability or versatility
If manual retraining process is used to update neural network, then network can adapt to new requirements, but the process requires human involvement and is not automated
Solution Approach 1:
The entire retraining workflow is automated through the control unit coordinating the selecting unit, annotating unit, and training unit. The system independently identifies when retraining is needed, selects appropriate training data based on actual processing failures, and executes the retraining process without requiring researcher or developer involvement, achieving full automation of the adaptation process.
Solution Approach 2:
The system proactively monitors neural network processing results and automatically initiates the retraining process when performance degradation is detected, before manual intervention would be required. The control unit is configured to trigger the automated workflow at predetermined time periods or when specific performance thresholds are met, enabling preemptive adaptation.
3Adaptability or versatility
If iterative retraining with updated data is performed to maintain optimal performance, then network can adapt to changing scenarios, but the process is complex and manual
Solution Approach 1:
The system merges multiple functions into a coordinated automated workflow: data selection based on processing failures, annotation acquisition, and iterative training are combined into a single integrated process triggered by the control unit. This unified approach simplifies the overall process complexity by automating the coordination between these previously separate manual steps.
Solution Approach 2:
The system autonomously manages the entire iterative retraining process, including automatically selecting which data to retrain on based on actual processing failures, acquiring the necessary annotated data, and executing the training iterations. This self-service capability eliminates the need for complex manual orchestration of multiple retraining steps.
Data Source
AI summary
The present disclosure provides a method and an apparatus for training a neural network and a computer server. The method includes: selecting automatically input data for which processing by the neural network fails, to obtain a set of data to be annotated; annotating the set of data to be annotated to obtain a new set of annotated data; acquiring a set of newly added annotated data containing the new set of annotated data, and determining a union of the set of newly added annotated data and a set of training sample data for training the neural network in a previous period as a set of training sample data for a current period; and training the neural network iteratively based on the set of training sample data for the current period, to obtain a neural network trained in the current period.