Dynamic Gradient Aggregation for Neural Network Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional distributed training methods treat all data sources equally, leading to inefficiencies when dealing with unbalanced or poor-quality training data, requiring significant manual filtering and tuning, which does not scale well with large datasets.
Innovation Solution
A computerized method for training a global model using Dynamic Gradient Aggregation (DGA), where gradients from each data set are weighted based on quality metrics to combine into a global gradient, automatically adjusting the model without manual filtering, and utilizing reinforcement learning to improve weight factor accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional distributed training methods treat all data sources equally, then the training process is simple to implement, but the training accuracy and convergence speed deteriorate when dealing with unbalanced or poor-quality data
Solution Approach 1:
The patent implements local quality by assigning different weight factors to gradients from different data sources based on their individual quality metrics. Each data source's gradient is weighted according to its specific quality characteristics rather than treating all gradients uniformly. This allows the system to give higher importance to high-quality data sources and lower importance to poor-quality sources, thereby improving training accuracy while automatically adapting to varying data qualities across sources.
Solution Approach 2:
The patent dynamically changes the weight factor parameters for different data sources based on computed quality metrics. The system calculates quality metrics for each data source and uses these metrics to adjust the weight factors in real-time during the training process. This parameter adaptation allows the training process to automatically respond to variations in data quality without requiring manual intervention or complex filtering procedures.
2Manufacturing precision
If manual filtering and sorting of training data sets is performed, then the training quality improves, but the time consumption and resource requirements increase significantly
Solution Approach 1:
The patent implements self-service by enabling the training system to automatically evaluate data source quality and adjust weight factors without requiring manual filtering or sorting. The quality metric computation and weight factor determination are performed autonomously by the system based on the characteristics of each data source, eliminating the need for human intervention in data preparation while maintaining high training quality.
Solution Approach 2:
The patent performs preliminary action by computing quality metrics for each data source before the main training process begins. This preliminary evaluation allows the system to pre-determine appropriate weight factors for each data source, so that during the actual training, the system can directly apply these weights without needing to perform filtering or sorting operations. This advance preparation reduces time consumption during the training phase.
3Reliability
If all training data sets are included in the training process, then the model robustness improves, but the negative impact of outlier values and poor-quality data increases
Solution Approach 1:
The patent addresses this contradiction by applying local quality assessment to each data source individually. The system computes quality metrics that reflect the specific characteristics and reliability of each data source, then uses these metrics to assign differential weight factors. This allows the model to learn from the diversity of multiple data sources (maintaining robustness) while simultaneously down-weighting or excluding the influence of poor-quality sources with outlier values (maintaining accuracy).
Solution Approach 2:
The patent introduces asymmetry in the treatment of different data sources by assigning non-uniform weight factors based on their quality metrics. Instead of symmetric treatment where all data sources contribute equally, the system creates an asymmetric weighting scheme where high-quality sources have greater influence and poor-quality sources have reduced influence. This asymmetric approach allows the model to benefit from diverse data sources while protecting against the negative effects of low-quality data.
Data Source
AI summary
The disclosure herein describes training a global model based on a plurality of data sets. The global model is applied to each data set of the plurality of data sets and a plurality of gradients is generated based on that application. At least one gradient quality metric is determined for each gradient of the plurality of gradients. Based on the determined gradient quality metrics of the plurality of gradients, a plurality of weight factors is calculated. The plurality of gradients is transformed into a plurality of weighted gradients based on the calculated plurality of weight factors and a global gradient is generated based on the plurality of weighted gradients. The global model is updated based on the global gradient, wherein the updated global model, when applied to a data set, performs a task based on the data set and provides model output based on performing the task.