Neural Network Training via Random Noise Injection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current neural network training methods lack diversity and robustness due to reliance on fixed training data sets, which can lead to overfitting and limited generalization capabilities, especially in supervised and unsupervised learning scenarios.

Innovation Solution

Introducing random data into the training process, including random input and output values for supervised learning, random variations in reward functions for reinforcement learning, and random input data for unsupervised learning, to enhance data diversity and adaptability.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If fixed training data sets are used, then training process is simple and efficient, but neural network lacks diversity and robustness

Engineering Contradiction:
ImproverobustnessVSAvoidtraining process complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent applies dynamics by transforming the static fixed training data set into a dynamic structure where random noise is continuously injected during training. The training process becomes adaptive and flexible, allowing the neural network to learn from varied data distributions rather than rigid fixed patterns, thereby improving robustness without requiring complete redesign of the training framework

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent changes the parameters of the training data by introducing random noise components. Instead of using fixed data points, the system modifies data parameters dynamically during training, adding variability to input data, target values, or both. This parameter transformation enables the network to handle diverse scenarios while maintaining training efficiency

Inventive Principle:
Principle #35Parameter changes

2Adaptability or versatility

If fixed training data sets are used, then training is efficient, but neural network exhibits overfitting and limited generalization

Engineering Contradiction:
Improvegeneralization capabilityVSAvoidtraining efficiency
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

The patent converts the harmful effect of overfitting into a beneficial learning opportunity by intentionally introducing random noise into the training data. This noise acts as a regularizer that prevents the network from memorizing fixed patterns, forcing it to learn more robust features that generalize better to unseen data while maintaining acceptable training efficiency

Inventive Principle:
Principle #22Blessing in disguise (Convert harm into benefit)

Solution Approach 2:

By making the training data dynamic through random noise injection, the system transforms the static training process into an adaptive one. The network learns from evolving data distributions, improving its generalization capability without sacrificing training efficiency, as the noise is injected during the training loop rather than requiring separate data generation steps

Inventive Principle:
Principle #15Dynamics

3Reliability

If random data is introduced into training, then data diversity and robustness improve, but training process becomes more complex

Engineering Contradiction:
ImproverobustnessVSAvoidtraining process simplicity
Core Design Contradiction:
ReliabilityVSEase of operation

Solution Approach 1:

The training system performs self-service by automatically generating and injecting random noise into the training data without requiring external intervention. The framework handles the complexity of noise generation, data augmentation, and training integration internally, allowing users to simply configure noise parameters while the system manages the entire process

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent merges the random noise injection mechanism with the existing training loop, combining data generation and training into a unified process. Instead of adding separate complex steps for noise generation and data preparation, the system integrates noise injection directly into the training cycle, maintaining simplicity while improving robustness

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentUS20240378434A1Method for training, optimization, and maintenance of neural networks via random noise injection.
Publication Date: 2024.11.14 MURPHY GABRIEL GOLDEN

AI summary

The methods disclosed herein introduce variability into neural network training and retraining by introducing random numbers into the training data sets for supervised or unsupervised learning modes, and by introducing random variability in the state transition probabilities and expected rewards for reinforcement learning.