Weighted Pattern Learning for Neural Network Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural network models struggle to accurately represent systems due to unequal reliability and accuracy of pattern data, as they inherently treat all data points equally during training, leading to suboptimal performance.
Innovation Solution
The method involves weighting pattern data based on its reliability, allowing more accurate patterns to contribute more significantly to the training process by modifying intermediate matrices in the least squares training method, enabling pattern-level weights to be used in neural network learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all pattern data is treated equally during neural network training, then the training process is simple and uniform, but the model accuracy deteriorates due to unequal reliability of different data points
Solution Approach 1:
The patent applies local quality by assigning different weights to different pattern data points based on their individual reliability characteristics. Instead of treating all data uniformly, each pattern receives a customized weight reflecting its quality, thereby improving model accuracy while managing complexity through targeted differentiation rather than universal complexity
Solution Approach 2:
The patent changes the parameter of data weighting from uniform to variable based on pattern reliability. By introducing a weight parameter that varies across different training patterns according to their reliability metrics, the system improves accuracy by emphasizing high-quality data while penalizing low-quality data, resolving the contradiction between simplicity and accuracy
2Reliability
If pattern-level weights are introduced to account for data reliability, then model performance improves, but the calculation complexity increases
Solution Approach 1:
The patent applies preliminary action by calculating pattern weights before the main training process based on reliability metrics. This pre-computation of weights allows the subsequent training to proceed efficiently with the weights already determined, avoiding the need for complex real-time calculations during training while still achieving improved model performance
Solution Approach 2:
The patent introduces an intermediary weighting mechanism that mediates between raw training data and the neural network training process. This weight layer acts as a bridge, translating reliability assessments into quantitative adjustments that improve model performance without directly complicating the core training algorithm
Data Source
AI summary
A method of training a neural net includes receiving a plurality of sets of data, each set representative of a plurality of inputs to the neural net and a resulting at least one output from the neural net and calculating a plurality of network weights for the neural network based on the received plurality of sets of data. Calculating the plurality of network weights including attributing greater weight in the calculation to at least one set of the plurality of sets of data than at least one other set of the plurality of sets of data.