QUANTUM INFORMATION AND MACHINE LEARNING:
COMMON APPROACHES AND TOOLS
[Project No. 20TTATQTa003 funded by SCS of Armenia]
Two broad fields of knowledge have recently experienced explosive growth: Quantum Information and Machine Learning.
The former relies on the fundamental principles of quantum physics and it aims to uncover the nature of information
and pave the way for quantum technologies: quantum computations, simulations, communications and metrology.
The latter combines Bayesian and frequentist statistics with artificial intelligence to analyze large data
and make better decisions in many areas of human activities, ranging from computer vision and self-driving vehicles,
to speech and text recognition and interactive human-machine interfaces.
Remarkably, both fields are based on probabilistic theories, and draw their advantages from
efficient implementation of computation and/or inference tasks. However, their interrelations
remain largely unexplored, and in this project we take several steps in the direction of uncovering them.
Thus, one of the main tasks in artificial intelligence is the search in a large database, or
optimization of a multi-parameter function. The Grover algorithm is a hallmark of quantum
computation that quadratically speeds up the search in an unstructured database. We suggest that
this search can proceed even faster, if the physical model of the database is subject to an external
open-loop control. The motivation for studying this problem is two-fold. First, in a hybrid system,
the quantum search can be embedded into a classical machine learning to greatly improve its speed.
Second, quantum algorithms rely on entanglement and superposition and uncovering their classical
counterparts can lead to new tools in classical data processing. Quantum computations suffer from
decoherence and noise affecting the qubits and quantum gates. We shall thus analyze various
methods for noise mitigation, and optimize them via the classical method of extrapolation to zero
noise limit. Another important example of problems relate to quasiprobability distributions that are
frequently employed in quantum optics for representations of continuous quantum variables. They
are more flexible than the usual probabilities and importing them to classical machine learning may
greatly assist supervised algorithms. We also envisage physics-inspired methods for unsupervised learning.
In summary, the main purpose of this project is to study pertinent aspects of quantum information and computation and to adapt the thereby learned quantum methods and tools to basic machine learning tasks.
Group meetings: Every Tuesday 13:30-15:00 at AANL
Presentations of the project results