ChemicalQDevice June 2023 R&D Update PDF 6/1/23.

ChemicalQDevice performed several Neuroimage Classifications that were based on a paper by Mari, A., et al. titled "Transfer learning in hybrid classical-quantum neural networks". A Pennylane hybrid machine learning model along with supplementary materials were incorporated to develop a medical image analysis framework. Using Google Kaggle brain tumor MRI datasets, the number of classes were increased from 2-Class to 4-Class to 10-Class. Direct comparisons between 10 classes of brain images have not been identified in literature, as detailed in MVP Iteration 7.

The key factors that improved model performance metrics of loss and accuracy were: 1) Several revisions to the quantum algorithm circuit size and layout of quantum gates 2) Batch size optimization, and 3) Other hyperparameters: step, step size, and threads. In addition, the Iteration 9 10-Class training loss/accuracy results of 0.115/0.969 are under review. Further research into whether quantum advantage was gained in the application is being conducted.

The Startup desires to evaluate early adopting customer projects to incorporate quantum machine learning (qml) algorithms into existing medical image analyses. The following three phases describe expected performance benefits as classical-quantum hybrid technology improves.

Phase I: Continuous implementation of linear quantum algorithms, but with higher grade quantum simulators for better performance and more complex problems. This includes different types of medical image applications, more classes, and larger datasets; with the startup seeking funding and talented medical/quantum individuals.

Phase II: A greater reliance on larger parameterized and interconnected linear quantum algorithms to further increase quantum advantage in applications, as simulators emerge. An option for emerging quantum computer applications also exists at this stage.

Phase III: Trainable Quantum neural networks are expected to be available for development on next generations of quantum simulators and quantum computers for greatest advantage. Medical images represented by quantum data derived from quantum sensors with emerging quantum memory may also present opportunities for algorithm development.

A preference towards quantum simulators will likely remain for some time due to 1) Noise free computing for mainstream regulatory acceptance, and 2) Evidence that quantum simulators will scale in numbers of qubits in gpu or multi-gpu environments. In several cases, quantum simulators running on cpus or gpus have been documented to have higher performance and speed than quantum computers while maintaining fidelity, as shown in Iteration 9 references. 20 pages attached, model performance is under review.

Additional References:
Quantum Embeddings by Aroosa Ijaz
QML: 2023 and Beyond by Maria Schuld