Entry for Saarah Nazar
Task V
The benefits of quantum machine learning are hidden but numerous. To begin, the inherent ability of quantum computers to process large amounts of data in parallel grants quantum machine learning a time advantage, as well as less tendency for bias to certain parts of the data provided (which would exist in a sequential model). In addition, quantum entanglement creates a way for kernel functions to detect nuance in datasets that classical kernels do not have access to. Transformations that are conducted on matrices of qubits representing the data can consist of conditional links between portions of the data, creating a data-driven algorithm for detection generated by the machine learning model. This is by far the feature (no pun intended) of quantum machine learning with the most potential. This can be demonstrated by a hypothetical algorithm for detecting malignancy of lung tumors that I generated using a quantum transfer learning model that detects commonly overlooked features such as adhesive single lung nodules (which are commonly misdetected), spinous protrusions’ external morphologies (textures are more easily detected by the conditional relation-based transformations allowed by quantum machine learning), and general distribution densities of nodules (uniform for benign versus cell nerosis and liquefaction for malignant). My original approach at MIT Beaver Works Summer Institute (bit.ly/sn-qcnn, https://github.com/saanaz379/lung-tumor-classifier) solely consisted of a quantum convolutional neural network. However, after much deliberation and research, I realized that the ideal approach would be a quantum transfer learning approach, utilizing high-quality XML files and images, such as the ones used by NoduleX, which would then run through a quantum convolutional neural network (QCNN), and then a quantum random forest to optimize the results of the whole model. The current best classical approach would be to simply use a Laplacian function, which is known for its antinoise ability. The approach’s accuracy would not even come close to the quantum approach’s accuracy, a crucial measure, especially for this model’s purpose.
The next steps in the field to bring it out of its infancy definitely include implementing and publishing implementations into the public domain of algorithms such as quantum random forests. This will definitely require greater hardware capabilities of both real quantum computers and simulators, both by increasing the number of qubits in a single computer, as well as exploring different hardware models, such as photonics. Another step would be to implement quantum optimization algorithms such as Quantum Approximate Optimization Algorithm and Quantum Kernel Alignment into existing and widely-used quantum machine learning libraries and frameworks such as Pennylane.
Pennylane is the quantum machine learning library I am most familiar with. This software provides a way to interface between quantum computing programming languages such as Qiskit and classical machine learning libraries such as TensorFlow and PyTorch. I believe the next big step for Xanadu (the company that created Pennylane) would be to start interfacing with lambeq, the Natural Language Processing library created by the legendary Bob Coecke, in order to allow quantum machine learning to break into the NLP domain.
The most profound applications of quantum machine learning would be the applications in optimizing the nuclear fuel cycle in order to grow the nuclear energy industry. For example, quantum machine learning could be used in comparing fuel cycle matrices (created by systems such as the Verifiable Fuel Cycle Simulation model (VISION) https://inldigitallibrary.inl.gov/sites/sti/sti/5094600.pdf to determine the accurate concentrations of Thorium and to minimize the toxicity of the waste while also minimizing Plutonium use (the more weaponizable component in the Thorium-Plutonium fuel cycle, as opposed to the Uranium fuel cycle. It can also be used to assess and address the uncertainties of neutronic parameters of a simple model of an ADS and determine the various fuel diffusion properties and constants and find the dispersion of various compounds such as Th O2, (Th, U) O2, and (Th, Ce)O2. Lastly it can help with the determination of neutron energy distributions to standardize the process utilizing variational quantum eigensolvers (VQE). These applications would give Thorium the advantage to be able to compete with Uranium in the current economic environment and create a safer world both in the short (avoiding nuclear disaster) and long term (preventing climate change) while maximizing energy production.