Neural Network Training via Task-Level Loss Weighting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network training methods focus attention at the sample level, ignoring the importance of different tasks, which can lead to imbalanced training and reduced recognition performance due to unbalanced sample proportions.
Innovation Solution
The method evaluates importance at the task level, adjusting loss function weights based on task-specific loss function values to prioritize important tasks during training, thereby improving network accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If sample-level attention is used in neural network training, then training simplicity is maintained, but task importance differentiation is lost leading to imbalanced training
Solution Approach 1:
The patent segments the loss function into multiple task-specific loss functions, each weighted according to its importance. This allows differential attention to be given to different tasks within the same sample, resolving the contradiction by maintaining operational simplicity while achieving training balance through loss function decomposition and weighted combination.
Solution Approach 2:
The patent applies local quality by assigning different weights to different task losses based on their specific importance. Instead of uniform treatment, each task's loss function receives localized attention through adaptive weighting, enabling the system to maintain simplicity while achieving balanced training across diverse tasks.
2Device complexity
If uniform loss function weights are applied to all tasks, then training process is simple, but recognition performance deteriorates due to unbalanced sample proportions
Solution Approach 1:
The patent introduces dynamic weighting of loss functions based on task importance and sample characteristics. The weights are not fixed but adapt during training, allowing the system to maintain simplicity while improving recognition performance through dynamic adjustment of training focus based on observed sample proportions and task difficulties.
Solution Approach 2:
The patent changes the parameters of the loss function by introducing task-specific weights that are adjusted during training. This parameter modification enables the training process to adapt to unbalanced sample proportions, improving recognition performance without significantly increasing overall process complexity.
3Measurement precision
If task-level importance evaluation is implemented, then recognition accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent performs preliminary evaluation of task importance before the main training process, establishing weight coefficients in advance. This preliminary action reduces computational complexity during actual training by pre-determining the relative importance of different tasks, allowing the system to achieve high recognition accuracy without excessive computational overhead during iterative training.
Solution Approach 2:
The patent introduces an intermediary mechanism (task importance evaluator) that bridges the gap between raw loss values and final weighted combination. This intermediary component simplifies the computational process by providing a structured method for evaluating and comparing task importances, reducing overall computational complexity while maintaining high recognition accuracy.
Data Source
AI summary
The invention provides a neural network training and application method, device and storage medium. The training method comprises: an obtaining step of obtaining a processing result and a loss function value of the processing result for at least one task after a sample image is processed in a neural network; wherein the neural network comprises at least one network structure; a determination step of determining importance of the processing result thereof based on the obtained loss function value; an adjustment step of adjusting a weight of the loss function for obtaining the loss function value based on the determined importance; and an update step of updating the neural network according to the loss function after the weight is adjusted.