Quantum Neural Network Training With Augmentation Functions

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The inefficiency and prolonged calculation time in adjusting parameters for quantum machine learning due to barren plateaus and increased resource demands as quantum circuits deepen, particularly in noisy intermediate-scale quantum computers, hinder effective training of quantum neural networks.

Innovation Solution

A training device and method that utilizes augmentation functions to connect shallow quantum circuits across multiple layers, optimizing weights without updating parameters directly in the quantum circuits, thereby reducing the need for extensive quantum computations and minimizing calculation time.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If quantum circuits are deepened to improve model capacity and representational power, then the ability to solve complex problems is enhanced, but calculation time increases significantly and training becomes inefficient due to barren plateaus

Engineering Contradiction:
Improvemodel capacityVSAvoidcalculation time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent segments the quantum neural network into multiple layers with alternating quantum and classical processing stages. Each layer processes part of the computation, allowing the model to achieve deep structure capabilities while breaking down the overall calculation into manageable segments that can be executed more efficiently.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Instead of directly optimizing quantum circuit parameters which leads to barren plateaus, the patent inverts the approach by optimizing classical parameters in the augmentation functions that connect quantum layers. This indirect optimization approach avoids the barren plateau problem while still achieving effective model training.

Inventive Principle:
Principle #13The other way round (Inversion)

2Adaptability or versatility

If quantum circuits are deepened to process complex information, then problem-solving capability is improved, but resource demands increase and training efficiency decreases

Engineering Contradiction:
Improveproblem-solving capabilityVSAvoidtraining efficiency
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

The patent introduces classical augmentation functions as intermediary components between quantum layers. These augmentation functions act as mediators that process quantum outputs and generate inputs for subsequent quantum layers, reducing the direct computational burden on quantum circuits while maintaining problem-solving capability.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The computation is segmented into discrete quantum and classical stages across multiple layers. Each layer handles a specific portion of the computation, allowing for more efficient resource utilization and faster training compared to deeply nested quantum circuits alone.

Inventive Principle:
Principle #1Segmentation

3Reliability

If parameters in quantum circuits are adjusted through quantum machine learning, then model performance is improved, but the number of quantum computations required increases

Engineering Contradiction:
Improvemodel performanceVSAvoidnumber of quantum computations
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent inverts the parameter optimization strategy by keeping quantum circuit parameters fixed and instead optimizing classical parameters in the augmentation functions. This approach achieves model performance improvement while significantly reducing the number of quantum computations required for training.

Inventive Principle:
Principle #13The other way round (Inversion)

Solution Approach 2:

The patent creates a hybrid architecture where classical functions copy and extend the capabilities of quantum circuits. The classical augmentation functions replicate and process information from quantum layers, allowing model performance to improve without proportionally increasing quantum computation requirements.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4679336A1Training program, training device, and training method
Publication Date: 2026.01.14 FUJITSU LTD
  • EP4679336A1 patent drawingFigure 1
  • EP4679336A1 patent drawingFigure 2~3
  • EP4679336A1 patent drawingFigure 4~5

AI summary

A computer acquires a training data set of a machine learning model. The machine learning model includes a quantum circuit in each of a plurality of layers and a generating function that generates, from an output of a first quantum circuit in a preceding layer in two consecutive layers, an input of a second quantum circuit in a subsequent layer. The computer determines a value of a parameter included in the generating function by training the machine learning model using the quantum computer that executes calculation of the quantum circuit in each of the plurality of layers and the training data set. The computer generates the trained machine learning model by setting the value of the parameter in the generating function.