Machine learning has long been part of our everyday lives. In its current form, it has nothing to do with quantum mechanics. This may change in the future, as new approaches that utilise the properties of quantum mechanical systems promise, for example, lower training costs - the main cost factor in classic deep neural networks. But what is one possible advantage of quantum mechanical systems?

Scaling of Hilbert space - The principle of superposition allows quantum mechanical neural networks to scale exponentially in their degrees of freedom, while the number of quantum subsystems only increases linearly. Even large language models with over 600 billion parameters could therefore, in principle, be executed on quantum hardware with just a few tens of qubits.

Entanglement - One property that can only be utilised on quantum mechanical systems is entanglement. While the benefits of entanglement are clearly evident in the Shor algorithm, for example, the question of the usefulness of this unique resource in quantum machine learning remains an open one at present.

Native processing of quantum signals - The research field of quantum sensing is one of the most mature branches of quantum high technologies and already has industrial applications. Current methods extract classical information from quantum sensors in order to process them; the richness of quantum information is lost in the process. Quantum machine learning enables the native evaluation of the sensors in order to utilise the full information content of the sensor technology.

At the CvO University of Oldenburg, research is being conducted into concepts for machine learning on quantum systems as hardware platforms. There are two main directions here:

Gate-based quantum computing: the aim here is to implement learning processes using quantum algorithms that can be executed on quantum computers. At the moment, nobody knows which approaches work best. For example, we are looking at what role the coding process of an input signal has on expressivity - the ability of the algorithm to generate as many linearly independent frequency components of the input signal as possible.

Quantum neural networks: This approach considers quantum systems that form a network through mutual coupling, somewhat by analogy with artificial neural networks. Signals can then be written into the quantum network, for example in the form of time series data through measurement processes. The complex dynamics of the network can then also be used for machine learning.

Contact:

Quantum Theory Group
Prof. Dr Christopher Gies