Neural Network Training via Confidence-Based Experience Replay

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing training methods for acoustic models based on neural networks are inefficient in terms of learning speed and training accuracy, particularly in speech recognition applications.

Innovation Solution

The proposed training method employs an experience replay mechanism, where samples with low confidence are selected more frequently for training, and well-recognized samples are gradually eliminated, enhancing the learning process through dynamic quota adjustments and domain-specific knowledge integration.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional training methods are used for acoustic models based on neural networks, then the training process is simple to implement, but the learning speed and training accuracy are insufficient

Engineering Contradiction:
Improvetraining accuracyVSAvoidtraining method complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent implements dynamic quota adjustment for training samples based on confidence levels. The quota function dynamically changes the frequency of sample selection during training iterations, transitioning from static uniform sampling to adaptive sampling that emphasizes difficult samples while reducing redundancy from easy samples.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent introduces feedback mechanisms where the model's confidence output is used to adjust training sample selection. The confidence level from each training iteration feeds back into the quota function, which then determines the sampling frequency for the next iteration, creating a closed-loop training process that continuously adapts to model performance.

Inventive Principle:
Principle #23Feedback

2Productivity

If all training samples are used with equal frequency, then the training process is straightforward, but the learning efficiency decreases due to redundant training on well-recognized samples

Engineering Contradiction:
Improvelearning efficiencyVSAvoidtraining process simplicity
Core Design Contradiction:
ProductivityVSEase of operation

Solution Approach 1:

The patent applies partial action by selectively training on subsets of samples based on confidence levels. Instead of uniformly processing all samples, the system focuses training capacity on samples that provide the most learning value (low confidence samples), while reducing or eliminating redundant processing of well-recognized samples (high confidence samples).

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The patent changes the sampling parameter (frequency) based on sample confidence levels. The quota function modifies the original uniform sampling distribution by assigning different weights to different samples, transforming the sampling strategy from a static equal-probability approach to a dynamic confidence-based approach.

Inventive Principle:
Principle #35Parameter changes

3Loss of time

If the training set includes all samples throughout training, then comprehensive coverage is achieved, but training time increases due to unnecessary processing of mastered samples

Engineering Contradiction:
Improvetraining timeVSAvoidtraining completeness
Core Design Contradiction:
Loss of timeVSReliability

Solution Approach 1:

The patent implements discarding of high-confidence samples from the training set as training progresses. Samples that achieve high confidence levels are gradually eliminated from the training pool, freeing up training time. The system recovers by reintroducing discarded samples only if their confidence degrades or if training requires review of previously mastered material.

Inventive Principle:
Principle #34Discarding and recovering

Data Source

PatentEP3379531B1Training method and apparatus for speech recognition
Publication Date: 2025.03.05 SAMSUNG ELECTRONICS CO LTD
  • EP3379531B1 patent drawingFigure 1
  • EP3379531B1 patent drawingFigure 2
  • EP3379531B1 patent drawingFigure 3

AI summary

A training method and apparatus for speech recognition is disclosed, where an example of the training method includes determining whether a current iteration for training a neural network is performed by an experience replay iteration using an experience replay set, selecting a sample from at least one of the experience replay set and a training set based on a result of the determining, and training the neural network based on the selected sample.