Quantum Machine Learning

【Topic 2: Quantum Intelligence.  Subtopic 2.4: Quantum Machine Learning】

I spend half a year to learn an online course “Quantum Machine Learning (QML)” from Prof. Peter Wittek of University of Toronto.  Peter, unfortunately, dies in an avalanche accident in Northern India around the time when the course ends.  From his course, I realize the efforts of researchers to make machine learning algorithms quantum-computable.  I am able to get my hands on real quantum computers, available from IBM, DWave and Rigetti.

For instance, after the course ends, I am able to run an algorithm called Quantum Approximate Optimization Algorithm (QAOA) by installing the most updated IBM Qiskit software version on my own computer and solving QAOA of a dozen max-cut nodes based on Wittek teaching of variational circuits.  In 2016, I am only able to run a few simple tutorial exercises on IBM Q-experience machines to get a feel of quantum computing.  After 3 years, IBM implements Qiskit software layer such that users can research on their own algorithms, not just those IBM tutorial exercises.  This is very powerful.

However, current quantum bits (qubits) from available processors are still too few.  A group of researchers, such as Caltech’s Prof. John Preskill, suggests a Noisy Intermediate Scale Quantum (NISQ) technology to resolve bigger problem with limited qubits.  Prof. Alan Aspuru-Guzik, an invited QML course speaker, also with University of Toronto, invents an eigen solver using NISQ and applies it to simulate chemistry molecular and material.  Alan is able to invent (1) new catalysts and new drugs, and (2) stronger substances that can help climate emission goal, for example.

Maria Schuld, also an invited QML course speaker, talks about her Kernel method of quantum machine learning.  I have her 2018 book “Supervised Learning with Quantum Computers”. People used to have wrong impression that the main purpose of quantum computer is to run quantum algorithms, not generating sample data.  Schuld’s kernel method can be used for preparing sample data (thus the term “quantum sampler”).  Specifically, the kernel method separates data with a “threshold hyperplane”, say into two groups, hence the problem becomes two-dimensional.  The kernel method can also separate data into infinitely-dimensioned spaces, such as an embedding space (say, Reproducing Kernel Hilbert Space, or RKHS).  Therefore, quantum data preparation is a big deal – the effort such as figuring out how quantum circuit is built may be more than that of the algorithm processing.

Most people when hearing the term Quantum Intelligence, they think it is just like here that the artificial intelligence algorithm is moved to the quantum computer to run, and the speed can be increased. However, in fact, quantum computers not only increase the speed of operation, but can perform calculations in accordance with the way the human brain thinks. The next subtopic has something to do in this respect.