Time, Cost, Efficiency: ResNet Quantum TL, TL Models PDF 8/13/23.

Goals and Objectives: Compare runtimes, costs, and efficiencies of two transfer learning methods across two datasets utilizing four different computing methods for insight to improve future quantum machine learning models.

Alternative Actions: Continue accessing hybrid computing resources on a per basis case, without detailed consideration of key metrics.

Stakeholders: Parties desiring to implement quantum computing methods that are now aware of CPU/TPU/GPU/Quantum financials.

Measurements: Utilizing Google Colab runtimes and computing method billing rates, run times and costs to evaluate efficiencies were primarily based on model type, number of classes, number of images, batch size, and whether a quantum algorithm was used.

Cost and Benefits: For 10 epochs, the CPU processing method was the most cost effective method across all model combinations ranging from $0.003-$0.044; and was the most efficient method for the quantum transfer learning model. TPU outperformed both GPU methods in the QTL model, but was not best in any specific category. V100 GPU was the fastest method for three out of four applications, ranging in times from 0.019 Hr-1.070 Hr; and was the most efficient method by a factor of two for TL 2 Class. A100 GPU had the greatest speed and efficiency for the larger 44 class classical model. V100 was relatively slow initializing the QTL model. A100 was the slowest to complete the first epoch in quantum models. There was no charge for accessing the quantum simulator. Accuracies overall favored QTL 2 Class, and TL 44 Class.

Prediction Outcome of Costs and Benefits over Time: Greater optimization of hybrid computing “talk” by industry experts will likely decrease both time and costs for greater efficiency in quantum models. In addition, improved fine tuning of CPU/GPU/Multi-GPU quantum simulators in cloud based compute platforms will also likely increase adoption rates and decrease costs.

Costs and Benefits Converted into a Common Currency: Quantum machine learning will likely offer fields such as medical informatics new utilities and higher accuracies due to processing datasets with qubits/qudits to gain advantage over conventional bits. Quantum simulators can likely accomplish these tasks, while emerging quantum computers with less noise may offer speed ups in the future. Prospective quantum machine learning applications for improving patient health are shared by leading healthcare organizations.