Adaptive Model Training for Low-Accuracy Image Tasks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Artificial intelligence models or deep learning models do not effectively enhance learning for training images with low identification accuracy, leading to reduced learning ability for more difficult tasks.
Innovation Solution
A model training method and system that adjusts sampling rates based on image identification rates, forcing the model to repeat learning on more difficult tasks by increasing the probability of selecting these tasks in subsequent iterations and reducing the probability of easier tasks, using a task filter to evaluate and adjust sampling rates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the model trains on all training data uniformly, then the training process is simple, but the model does not strengthen learning for training images with low identification accuracy
Solution Approach 1:
The patent implements a feedback mechanism where the model's identification accuracy on each sub-task is evaluated, and this evaluation result is used to dynamically adjust the sampling rate for subsequent training iterations. Specifically, sub-tasks with lower identification accuracy receive higher sampling rates, creating a closed-loop system that continuously adapts the training process based on performance feedback.
Solution Approach 2:
The training process transitions from a static uniform sampling approach to a dynamic adaptive sampling approach. The sampling rate for each sub-task is no longer fixed but changes over time based on the model's evolving identification accuracy, allowing the training process to automatically focus on difficult tasks as the model improves.
2Reliability
If the model focuses on difficult tasks by increasing their sampling rate, then learning ability for difficult tasks improves, but training time increases
Solution Approach 1:
The patent employs periodic evaluation and adjustment of sampling rates across multiple training iterations. Instead of continuously adjusting or using a fixed rate, the system periodically assesses identification accuracy and recalibrates sampling rates, creating a rhythm of evaluation-adjustment-train that balances focus on difficult tasks with overall training progress.
Solution Approach 2:
The sampling rate parameter is dynamically changed based on identification accuracy metrics. By adjusting this key parameter throughout training, the system optimizes the balance between focusing on difficult tasks and maintaining efficient overall training, rather than using a static parameter throughout the entire training process.
3Productivity
If uniform sampling rate is used for all training data, then training efficiency is maintained, but identification accuracy for different sub-tasks becomes unbalanced
Solution Approach 1:
The patent applies different sampling rates to different sub-tasks based on their specific identification accuracy levels. Instead of treating all training data uniformly, the system identifies which sub-tasks require more attention and applies localized adjustments to their sampling rates, ensuring each sub-task receives appropriate training intensity.
Data Source
AI summary
A model training method and a model training system are disclosed. The method includes the following. In a first iteration training, first training data corresponding to a first sub-task and second training data corresponding to a second sub-task are input to a target model. A first and a second image identification rate of the target model with respect to the first and the second sub-task are evaluated respectively according to a first output corresponding to the first sub-task and a second output corresponding to the second sub-task from the target model. A first and a second sampling rate respectively corresponding to the first and the second training data in a second iteration training are adjusted according to the first and the second image identification rate. The first and the second sampling rate are respectively negatively correlated to the first and the second image identification rate.