Quantum Annealing Performance Prediction via Bayesian Neural Network Ensemble
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Obtaining quantum annealing performance metrics to train machine learning models that predict performance outside the domain of the training data is challenging due to the scarcity of large quantum annealing hardware, leading to poor extrapolation by ML models.
Innovation Solution
A hybrid, SLA-oriented mechanism using a Bayesian Neural Network (BNN) trained once, with models sampled to form a dynamic ensemble to capture uncertainty in predictions, allowing for extrapolation beyond the training data domain and informing job placement decisions in computing infrastructures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If ML models are trained on limited quantum annealing hardware data, then the training process becomes feasible, but the model performance deteriorates when extrapolating to out-of-distribution values
Solution Approach 1:
The patent creates multiple synthetic copies of quantum annealing performance data by generating artificial datasets that mirror the statistical properties of real hardware performance. These synthetic datasets are created using physics-based models and simulations, allowing the ML model to be trained on abundant synthetic data while maintaining generalization to real hardware scenarios, thus resolving the contradiction between limited training data and reliable extrapolation.
Solution Approach 2:
The patent transforms the training approach by changing from using actual hardware performance parameters to using simulated parameters derived from physics models. By training on simulated parameters and then transforming predictions to physical parameters, the model can extrapolate reliably to out-of-distribution hardware configurations without requiring actual hardware data for each scenario.
2Quantity of substance
If more quantum annealing hardware is deployed to collect performance metrics, then the training data quantity increases, but the hardware cost and complexity increase
Solution Approach 1:
Instead of deploying more physical quantum annealing hardware to collect training data, the patent creates synthetic copies of performance data through physics-based simulations. These synthetic datasets replicate the statistical characteristics of real hardware performance across various configurations, providing abundant training data without requiring additional physical hardware infrastructure.
Solution Approach 2:
The patent introduces physics-based simulations as an intermediary between theoretical models and actual hardware performance. These simulations act as a mediator that generates training data by modeling the physical behavior of quantum annealers, eliminating the need for direct hardware deployment while still producing accurate performance metrics for training.
3Loss of time
If a single BNN is trained, then the training time is reduced, but the ability to capture uncertainty in predictions is compromised
Solution Approach 1:
The patent creates multiple instantiations (copies) of the BNN model, each trained on different subsets of synthetic data or with different hyperparameter configurations. These multiple models collectively capture uncertainty by providing a distribution of predictions, while each individual model remains computationally efficient. The ensemble of models provides both speed and uncertainty quantification.
Solution Approach 2:
Instead of training a single comprehensive model that would require extensive training time and data, the patent trains multiple simpler models on partial datasets. This partial action approach allows each model to train quickly while the collection of models collectively provides comprehensive uncertainty coverage through their combined predictions.
Data Source
AI summary
A method includes generating multiple instantiations of a machine learning model that has been trained, with a training dataset, to model a function that is operable to generate a prediction regarding a quantum process metric, sampling the instantiations of the machine learning model, populating an ensemble with the instantiations of the machine learning model obtained as a result of the sampling, and generating respective predictions, regarding the quantum process metric, with each of the instantiations of the machine learning model in the ensemble.