Two-Stage Quantum Autoencoder Training for Faster Optimization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The challenge in quantum machine learning is the difficulty in finding local solutions for noise removal in variational quantum algorithms, leading to prolonged training times and accuracy degradation in quantum autoencoders, especially when dealing with large problem sizes.

Innovation Solution

A two-stage machine learning approach is employed, first training a quantum autoencoder using input data with identical qubit states for a subtask, followed by training with different qubit states for the main task, to avoid local solutions and reduce training time.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional machine learning is used on quantum autoencoder with large problem sizes, then more variational parameters need to be optimized, but training time is prolonged and accuracy degradation occurs

Engineering Contradiction:
Improveaccuracy of quantum autoencoderVSAvoidtraining time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The training process is segmented into two distinct stages: first training with input data where all qubits are in the same state, then training with input data where qubits are in different states. This segmentation allows the model to learn basic patterns first before handling more complex variations, reducing overall training time while maintaining accuracy.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The first training stage serves as a preliminary action that prepares the quantum autoencoder by establishing foundational learning patterns with simplified input data. This preliminary training with identical qubit states creates a good initialization that accelerates subsequent training with more complex different-state inputs.

Inventive Principle:
Principle #10Preliminary action

2Reliability

If traditional machine learning is used on quantum autoencoder, then comprehensive training is attempted, but local solutions are difficult to find and optimization becomes challenging

Engineering Contradiction:
Improvequality of local solutionVSAvoidcomplexity of optimization process
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The optimization process is divided into two phases with different input characteristics. The first phase optimizes with identical qubit states which creates a simpler optimization landscape easier to navigate, while the second phase handles the full complexity. This segmentation makes the overall optimization more manageable and reliable.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The first training stage performs a preliminary optimization that establishes a good baseline solution. By initially training with identical qubit states, the system performs preliminary actions that set favorable initial conditions for the subsequent optimization with different-state inputs, making it easier to find high-quality local solutions.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4621650A1Machine learning program, machine learning method, and information processing device
Publication Date: 2025.09.24 FUJITSU LTD
  • EP4621650A1 patent drawingFigure 1
  • EP4621650A1 patent drawingFigure 2A~2B
  • EP4621650A1 patent drawingFigure 3A~3B

AI summary

A machine learning program including instructions which, when executed by a computer, cause the computer to execute processing including: executing first machine learning on a quantum autoencoder by using first input data and second input data in which states of the respective qubits are mutually the same; and executing second machine learning on the quantum autoencoder trained by the first machine learning by using third input data and fourth input data in which states of the respective qubits are mutually different.