Libby Heaney
7 min readDec 12, 2018


Exploring the unknown-unknowns: perspectives on quantum machine learning

What is quantum machine learning? How far away are we from tenable quantum machine learning systems? And what might quantum machine learning be used for?

As a former quantum physicist turned artist, I am interested in the impacts of emerging technology. This article therefore explores the known unknowns of quantum machine learning, as, at least to me, this area is really not clear at the moment. While I am not being particularly critical in this post, I hope to generate some discussion outside the quantum computing community about these new forms of technology and what they might mean for us and the earth.

After reviewing the literature, there seems to be three, or maybe four, options for a future quantum machine learning.

The first option is using current machine learning algorithms to spot statistic patterns in quantum data. What is quantum data? It is data produced by quantum physical systems. These are things like atoms, superconductors, lasers and molecules that need quantum mechanics for their physical description. If one collects data from these, current (non-quantum) machine learning techniques can analyse and interpret these quantum state-spaces up to a point, which is useful for the development of quantum technologies and new materials. However, quantum physical systems can produce counter-intuitive patterns. That is, patterns that cannot be produced by everyday classical systems. It is very unlikely current classical machine learning algorithms will be able to recognise quantum patterns. What we are talking about here is the ability to distinguish between patterns produced by large quantum physical systems like clusters of atoms and most other (classical) patterns like people walking in a city or the spread of viral content online.

For what is possible, the limitation will be the size of the neural networks you can run on existing super-computer clusters because the quantity of data from quantum systems grows at a much bigger rate than the data from non-quantum systems (even compared to really complex systems like earth’s climate).

This type of quantum machine learning system is already here, but in my mind it should not be considered quantum machine learning as no quantum computers are involved.

The second option is a ‘hybrid’ model, where some of a classical (non-quantum) machine learning algorithm is outsourced to a quantum computer. Here, the majority of the learning process takes place on a classical computer as it does now and only parts requiring a speed up are passed to a quantum computing chip. This means that bottle-neck routines on current supercomputers could be passed to a quantum computing chip and be run more efficiently. Using this hybrid method, simulations of quantum systems, that are currently challenging to complete, such as the simulations of large complex molecules can be substantially sped up. Quantum algorithms such as the 2008 HHL algorithm for solving linear equations will assist in providing the speed up. Scientists proposed this technique to solve the electronic states of molecules, which is useful for drug design and for new materials for energy storage and generation. This ‘hybrid’ technique may lead to more refined understandings of other complex systems such as our brain and the climate. The downsides of a hybrid approach are the time and efficiency costs of passing data to and from the quantum computer. Quantum RAM or qRAM may be needed. Suitable hardware is a hugely limiting factor. Non-general quantum computers could be used instead of universal quantum computers. These could be similar to D-Wave’s quantum computing system, which performs well only for specific tasks and are already available. As all outputs from the algorithm will be classical, it seem unlikely that it will spot patterns beyond classical machine learning algorithms.

The third option is to radically adapt and run existing machine learning algorithms on a quantum computer. I write ‘radically adapt’ as quantum computers run very differently compared to their classical counterparts. This is what some researcher have termed quantum machine learning (see for instance Xia and Kais Nature Communications 2018). This approach is similar to the hybrid model, except the whole learning procedure runs on a quantum computer. Again both classical and quantum data can be used for the training set. The benefits of using a quantum computer arises when the learning utilises a quantum speed up as in the ‘hybrid’ option. Since this approach adapts or rather quantizes existing machine learning algorithms, it is unclear whether it will be able to spot patterns not recognisable by classical computers, even though the output may be a quantum vector.

This approach has been theoretically proposed, but it is difficult to say when such systems will be put to use in efficient and cost effective ways. Hardware challenges are the main hurdle. Doing this depends on the ‘holy-grail’ development of a full scale quantum computer. While Google, IBM and co have been touting this possibility for a while now, it seems to me they are currently aiming for a ‘dirty’ quantum computer. A ‘dirty’ quantum computer runs a quantum algorithms badly on a large number of qubits, which means they will be extremely error prone. So far the biggest device, announced on 11th December by startup IonQ, apparently contains 160 error-prone qubits. Companies seem to be taking the dirty route because they are in a ‘space-age style race’ for ‘quantum supremacy’. That is, to create a quantum computer that computes something that is not possible on the biggest digital computer. However, it is likely years if not decades before something meaningful comes out.

With that said let us turn to my final option for a future quantum machine learning, which I’ll dub ‘genuine quantum machine learning’. Rather than simply speeding up existing algorithms, genuine quantum machine learning may ‘see’ the universe in completely different ways to how we do now. Given the nature of quantum mechanics, it is likely that a genuine quantum machine learning will access facets of reality that are seemingly off-limits to non-quantum systems, including us.

A very tentative glance in this direction was mentioned in Biamonte et al arxiv/1611.09347v2. There, the team argue that since quantum information processors can produce patterns that are computationally difficult to produce by classical computers, they should also be able to recognise patterns that are equally difficult to recognise classically.

As a practising artist, I would like to imagine what this might mean. I believe genuine quantum machine learning will explores the unknown-unknowns. When this occurs, we will be in a distant but plausible future where full scale quantum computers are up and running. New quantum algorithms will be produced and tested with ease.

Genuine quantum machine learning algorithms will be able to ‘see’ reality in fundamentally different ways to humans. It is doubtful whether our biological brains will be able to grasp the statistical patterns they find. And technically it may even be impossible to convert the multi-dimensional and exponentially large quantum outputs into standard bit strings that we could then work with.

Let us compare that to now. Currently, unsupervised classical machine learning algorithms can already recognise detailed statistical patterns humans are not able to see. Frequently humans do not understand how deep learning algorithms make their predictions. But given those predictions, it is likely someone will understand what the outputs mean once they are analysed and tested. I believe this will change — at least for a while we’ll be entering a quantum twilight zone.

While quantum patterns are purported to exists mostly at small length and time scales and at cold temperatures, scientists have shown they also exist in biological systems like photosynthesising algae. Biological systems are large, warm and wet so finding quantum patterns there is quite strange. Scientists also have theoretically shown that entropy lowering mechanisms may assist quantum patterns in manifesting at larger length scales and at higher temperatures. Therefore it is highly likely that genuine quantum machine learning will start to recognise quantum patterns in places scientists never expected to find them. Genuine quantum machine learning algorithms will be able to spot hidden quantum signatures embedded within classical data, perhaps showing evidence for parallel universes or theories of quantum gravity. Maybe dark matter and dark energy will come into focus. Quantum machine learning might provide new insights into individual and collective consciousness. It’s ability to spot and produce long range, non-local correlations suggests it may revolutionise natural language processing and other forms of communication synthesis and analysis. And it will undoubtedly revolutionise artistic and cultural understandings in ways we cannot imagine.




Libby Heaney

Artist, PhD quantum physicist, researcher/lecturer at the Royal College of Art: trying to make the inconceivable visible and questioning what this means.