Johannes Nokkala
TCSMT Postdoctoral Researcher
Department of Physics and Astronomy
There are many tasks of practical interest that are not suited for conventional computer algorithms. Suppose for instance that we want to turn speech into text. Although it is easy for us to prepare examples of speech and the corresponding text, we don’t know how to translate what our brains do into, say, if-then statements. To tackle such problems, we need algorithms that can infer the unknown transformation from examples. This is known as supervised machine learning. The basic idea is to use an input-output system that can transform the input in a wide variety of ways depending on how we choose its parameters, and then adjust the parameters to make it mimic the examples. Provided that the true unknown transformation is within its reach, the hope is that with enough examples we can optimize the parameters in such a way that the system continues to perform well even when given new inputs that were not included as examples.
Since I’m a physicist it might be asked why I’m interested in something that sounds more like computer science than physics. My fascination is in fact related to harnessing the dynamics of physical systems for machine learning tasks. The basic idea is almost the same as before, with the difference that the inputs are used to drive the dynamics of the system and the outputs are adjustable functions of its response. Importantly, if the system is complex enough then even very simple functions of the response may be enough to solve nontrivial tasks. Furthermore, generic complex systems are often suitable, opening the way to directly harvest information processing from physical systems that can be controlled. This should be contrasted with conventional computers that can be quite powerful but require precise engineering.
The flexibility to use almost any controllable system makes the direct approach particularly appealing when going beyond classical physical systems to the quantum realm, where the quantum counterparts of conventional computers are still quite far from reaching their full potential due to issues with, e.g., scalability and noise. Indeed, at this point when there are already controllable small to intermediate scale quantum systems it is of great interest to consider what they can be used for.
I was first introduced to the topic two years ago when I was starting a postdoc at IFISC in Palma, Spain. If I succeeded to pique your curiosity about it, I recommend taking a look at our most recent work, currently available in arXiv, which is a perspective article providing more details and background as well as an overview of proposals, experiments and hurdles. I hope to continue pursuing this exciting research avenue during my time here in Turku.
Johannes Nokkala
TCSMT Postdoctoral Researcher
Department of Physics and Astronomy
