Neural Network Training Method Using Reference Hidden Node Exclusion

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current neural network training methods face challenges in efficiently training recurrent neural networks for sequential data recognition, particularly in maintaining long-term memory values and preventing overfitting, which affects the network's ability to generalize and classify new input patterns effectively.

Innovation Solution

The method involves selecting a reference hidden node in each time interval, excluding it from the training process, and maintaining its long-term memory value across intervals, while connecting remaining hidden nodes across different time intervals, thereby updating connection weights based on the training results to enhance generalization and reduce overfitting.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If all hidden nodes are trained using standard backpropagation, then the network learns patterns effectively, but the training time increases and overfitting occurs

Engineering Contradiction:
Improverecognition accuracyVSAvoidtraining time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent extracts and excludes a reference hidden node from the training process while maintaining its long-term memory value. This selective exclusion reduces the computational burden on the reference node, thereby reducing overall training time while preserving the network's recognition accuracy through the node's maintained memory function.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent segments the hidden layer into remaining hidden nodes and a reference hidden node with distinct functional roles. The reference node maintains long-term memory without standard training, while remaining nodes undergo normal training. This segmentation allows parallel processing paths that reduce training time while maintaining recognition accuracy.

Inventive Principle:
Principle #1Segmentation

2Adaptability or versatility

If standard training is applied to all hidden nodes, then the network adapts to training data, but overfitting occurs and generalization ability decreases

Engineering Contradiction:
Improvelearning capabilityVSAvoidgeneralization ability
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

By extracting the reference hidden node from standard training and maintaining its value across time intervals, the patent prevents this node from overfitting to training data patterns. This extracted node serves as a stabilizing element that improves generalization ability while the remaining nodes continue to learn from training data.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent changes the training parameter for the reference hidden node by excluding it from standard backpropagation updates. This parameter change (from trained to maintained) creates a regularization effect that prevents overfitting and enhances the network's ability to generalize to new data.

Inventive Principle:
Principle #35Parameter changes

3Reliability

If the reference hidden node is excluded from training, then overfitting is reduced, but the node's contribution to pattern recognition may be limited

Engineering Contradiction:
Improvegeneralization abilityVSAvoidrecognition accuracy
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The reference hidden node serves multiple functions: it maintains long-term memory values across time intervals, provides regularization to prevent overfitting, and still contributes to pattern recognition through its maintained values. This multi-functionality ensures that exclusion from training does not limit its recognition contribution.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The reference hidden node maintains continuous useful action by preserving its value across time intervals without being reset or updated through training. This continuity allows the node to contribute consistently to pattern recognition while preventing overfitting, balancing both generalization and accuracy.

Inventive Principle:
Principle #20Continuity of useful action

Data Source

PatentEP3059699B1Neural network training method and recognition apparatus
Publication Date: 2020.02.26 SAMSUNG ELECTRONICS CO LTD
  • EP3059699B1 patent drawingFigure 1
  • EP3059699B1 patent drawingFigure 2
  • EP3059699B1 patent drawingFigure 3

AI summary

Disclosed is a neural network training method and apparatus, and recognition method and apparatus. The neural network training apparatus receives data and train a neural network based on remaining hidden nodes obtained by excluding a reference hidden node from hidden nodes included in the neural network, wherein the reference hidden node maintains a value in a previous time interval until a subsequent time interval.