Computers are slowly turning human. No, not literally. We will not have to bow down to our silicon overlords anytime soon. They are, however, starting to ‘learn’ the same way we do ourselves. With the advent of superfast quantum computers, this development is gaining momentum. Vedran Dunjko develops quantum algorithms and techniques that could put this principle of ‘machine learning’ into overdrive.
How to recognise a chair
Instead of needing human-prescribed rules to reach the desired result, computers are starting to think for themselves. Take image recognition: computers today are able to discern particular objects in pictures. They can, for instance, recognise chairs. Not because humans have told them about legs, seats and backrests, but because their algorithms have been fed millions of pictures of chairs out of which they distilled the basic features that qualify an object as being a chair. This principle of ‘machine learning’ is already being widely used in the modern world and is taking up ever-increasing amounts of computational power. Computer scientist and theoretical physicist Vedran Dunjko is working on ways to harness the power of the enigmatic quantum computer to satisfy the needs of these new applications and to take machine learning to the next level.
A one with twenty-seven zeros
‘In 2008 and 2014, ideas were put forward that showed that quantum computers can perform some very important data processing tasks exponentially faster than conventional computers,’ says Dunjko. ‘They can perform tasks like extracting the global shape of data – which can be used in all sorts of problems from classifying chairs to understanding large proteins – in such a way that the computational effort is only logarithmic. Such analyses are in general extremely demanding. If you were to analyse data with just 100 points, for instance, you would easily end up with having to do a billion billion billion operations – a 1 followed by twenty-seven 0s. A classical computer would never be able to process this amount of information, while a quantum computer could accomplish such a feat by performing just millions of operations. This can be a game-changer.’
Quantum computers derive their power from the unit of information they use. Conventional computers use ‘bits’, which are essentially 0s and 1s. Quantum computers use ‘qubits’, which rely on the properties of small elementary particles like photons (light particles) and can counterintuitively take on the values of 0 and 1 at the same time. This strange quantum phenomenon, which in reality is quite a bit more complicated than the previous sentence would suggest, potentially translates into a mind-boggling increase in computing power.
The more bits the better
Just as with conventional computers, the adage is: the more bits the better. For quantum computers to reach their full potential, they may need millions of qubits. These machines will not be available in the near future. The current record holders are quantum computers made by IBM and Google that work with no more than 53 qubits. This is why Dunjko and his research team are developing machine learning algorithms and techniques for earlier, limited quantum computers called Noisy Intermediate-Scale Quantum (NISQ) computers that operate with around hundreds or thousands of qubits.
‘Working on quantumcomputers is like a forgotten dream. There is no doubt that this is what I should be doing.’
Back to the old mindset
‘Honestly speaking, these are just bad quantum computers,’ Dunjko says with a chuckle. ‘They are limited in size and prone to error. But we already know they can leave classical computers in the dust regarding some exotic computations. We “just” need to make them useful. The main problem is that until recently we approached algorithm creation for these machines in the same way as we do with today’s fully developed conventional computers, in which there are no errors and almost no size restrictions. We need to go back to the mindset of the 1940s, a time when you would have been called insane if you had been considering machines that work with billions of bits. Machine learning applications show promise of still being able to function even under the restrictions that NISQ computers impose.’
As is often the case with fundamental advances in science, it is difficult to pinpoint the first practical application of NISQ computers, although, according to Dunjko, most researchers in the field agree on the best bet. ‘The first clear advantage of these machines over conventional computers will probably be in the area of predicting what happens in quantum systems – the realm of atoms and molecules,’ he says. ‘Here, you have a nice match between the object of research and the way quantum computers are built. Think, for instance, about the creation of molecules with certain predetermined properties, benefiting drug design. I believe quantum machine learning can help us do this better.’
The quantum computer of tomorrow
Looking into the future, Dunjko hopes to produce the first non-controversial example of an advantage of a real-world quantum computer over a conventional one. ‘This is one of the holy grails in our research field,’ he says. ‘And in twenty to thirty years I hope we can put the NISQ era behind us and have universal quantum computers. At the moment, though, we feel it’s extremely important to make use of the machines that are available in the short term. These stimulate research, yield new insights into quantum technologies and raise the confidence of everyone involved. This way, NISQ computers pave the way towards the full-blown quantum computer of tomorrow.’
Vedran Dunjko (Zagreb, 1980) started thinking about quantum computing and machine learning when he was in first grade at high school. His interest was triggered by his older brother, who at that time had already become a physicist. ‘I tried reading about quantum computing and artificial intelligence. Both topics were equally fascinating… and totally beyond me,’ he says. After some forays into mathematics and biology, Dunjko has now arrived at the Leiden Institute of Advanced Computer Science (LIACS) where he is an assistant professor conducting research at the intersection of computer science and quantum physics. ‘In some sense, this is like a forgotten dream,’ Dunjko says. ‘There is no doubt that this is what I should be doing.’