Quantum Neural Network Pre-Training for Reduced Training Time
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The long training time required for quantum neural networks (QNNs) due to the need for many measurements in each optimization step, which can erode the quantum advantage in solving problems like image classification.
Innovation Solution
The implementation of pre-training for quantum neural networks, where a first QNN is trained using a dataset different from the second dataset, and the pre-trained setting is used to generate outputs for the second QNN, which is then trained in a fixed setting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the entire QNN is trained from scratch using many measurements in each optimization step, then the QNN can be trained to solve problems, but the training time becomes excessively long and erodes quantum advantage
Solution Approach 1:
The patent applies pre-training by first training a first QNN on a first dataset to establish an initial set of weights and configurations. This preliminary training creates a foundation that is then transferred to a second QNN, which only needs to be trained on a second dataset using the pre-trained settings as initialization. This preliminary action significantly reduces the training time required for the second QNN while maintaining training effectiveness.
Solution Approach 2:
The training process is segmented into two distinct phases: a first training phase that trains a first QNN on a first dataset, and a second training phase that trains a second QNN on a second dataset using the pre-trained settings from the first phase. This segmentation allows each phase to focus on specific tasks, with the first phase establishing general features and the second phase refining them for specific problems, thereby reducing overall training time.
2Loss of time
If a first QNN is trained on a first dataset and then used to generate outputs for training a second QNN on a second dataset, then training time is reduced, but the complexity of the training process increases
Solution Approach 1:
The pre-trained settings from the first QNN act as an intermediary that bridges the two training processes. Instead of training the second QNN from scratch, the pre-trained settings serve as a transferable intermediary that encapsulates useful features learned during the first training phase. This intermediary reduces the computational burden of the second training phase while organizing the complexity into a manageable transfer learning framework.
Solution Approach 2:
The first QNN's pre-trained settings are designed to be universally applicable to the second QNN's training process. The same pre-trained weights and configurations can be transferred across different datasets and tasks, making the training process more efficient and reducing the need for complete retraining. This multi-functionality of the pre-trained model reduces training time while the modular architecture manages complexity.
Data Source
AI summary
A method includes training a first QNN by sending a first dataset into the first QNN to generate a first output and configuring the first QNN into a first setting based on the training. The method also includes receiving a second dataset, using at least a portion of the first QNN to generate a second output using the first setting, and sending the second output to a second QNN, operatively coupled to the first QNN, to train the second QNN. The second QNN is configured in a fixed setting during training of the first QNN.