Ellis Fellows program Quantum and Physics based machine learning (QPhML)

Increased computational power and data availability, as well as algorithmic advances, have led to impressive machine learning applications in many areas such as computer vision, pattern recognition, robotics and AI. At the same time, conventional CMOS technology is reaching its physical limits and the energy consumption of computing is reaching alarming proportions. There is therefore a great need to design novel computing paradigms that face these challenges.

The aim of the Ellis program Quantum and Physics based machine learning (QPhML) is to use concepts from quantum physics and statistical physics to develop novel machine learning algorithms with the ultimate aim to realize novel future, possibly energy efficient, hardware implementations.

The program is part of the recent European initiative called Ellis (European Laboratory for Learning and Intelligent Systems) ELLIS to stimulate research on machine learning by building networks of top reseach groups in Europe.

The program focuses on the following directions:

Quantum enhanced machine learning.

Quantum devices are nearing the noisy intermediate scale quantum (NISQ) era, corresponding to machines with 50 to 100 qubits and capable of executing circuits with depths on the order of thousands of elementary two qubit operations. NISQ devices may provide computational advantages over classical supercomputers for various machine learning problems, which includes sampling from hard-to-simulate probability distributions for Bayesian methods and the Quantum Boltzmann Machine and linear algebra problems (for instance for kernel methods or deep learning). It is hoped that the application of NISQ technology to machine learning may be one of the first instances exhibiting genuine quantum advantages.

Statistical physics approach to machine learning

Noise plays a fundamental role for learning in large neural networks. Rather than designing reliable bits and use software to generate random numbers, an appealing alternative is to design hardware that is noisy by design. Such devices would be much more energy efficient. Methods from non-equilibrium statistical physics are well suited to improve our understanding of stochastic systems. An example is the use of physically coupled replicas that have been shown to be very effective for hard combinatoric or strongly non-linear learning problems. In addition, the observation that physical replicas resemble Trotterized quantum systems provides a promising new research direction for the design of stochastic or quantum learning algorithms. Another link between quantum and stochastic systems is the observation that sign free quantum systems can be mapped onto classical stochastic diffusion problems.

Using machine learning for quantum physics

The challenge of quantum many-body physics is to efficiently describe and control exponential numbers of parameters of quantum systems. Better characterization of such systems will lead to the understanding of quantum materials such as high-temperature superconductors or topological insulators. Enhanced control of immense parameter spaces will improve the understanding and design of quantum devices, enabling quantum computers and networks. For this problem, machine learning offers a new option.

The Ellis program is realized through the active involvement of the Ellis Fellows (senior researchers) and Ellis Scholars (more junior researchers). In addition, the Ellis Guests, are affiliated researchers.





Ellis QPhML workshop. The DALI conference. San Sebastian, 5 September 2019