Besides chemistry and cryptography, one potential application of quantum computing that’s generating significant excitement is machine learning. Here, Jungsang Kim, who leads the Multifunctional Integrated Systems Technology group at Duke University in Durham, North Carolina, and co-founded quantum-computing company IonQ, based in College Park, Maryland, explains why.
What’s so interesting about quantum machine learning?
The conventional approach to machine learning works well for many applications, but when you try to increase the accuracy or make the model more robust, classical machine learning tends to get expensive very quickly. You can be happy with 70% accuracy when recommending the next song in a playlist or running an advertisement, but for more-crucial activities, such as autonomous driving, the accuracy needs to be much better. Customers want to see whether there are fresh solutions, and quantum machine learning might be just that.
What does quantum computing bring to machine learning?
In classical deep learning, you don’t know what the data structure is beforehand. So, you come up with a model, and if it doesn’t fit the data structure properly, you add more parameters to the model and use a lot more data to fit it, and hopefully things converge. Although this has proved to be very powerful, the model gets very complicated and expensive to train quickly. But using the mathematical structure behind quantum descriptions of nature, called Hilbert space, you can describe certain complex structures with very few parameters. These quantum models can be much more ‘expressive’ than their classical counterparts, and can capture complex patterns in data that are otherwise difficult to capture. This is not all that surprising, really: if you consider highly correlated quantum materials in which the electrons interact strongly, they can have a very simple quantum interaction that is very complex to describe classically.
Will quantum machine learning be competitive with classical machine learning?
We speculate that these more-expressive quantum models will be easier to train, and early research results suggest as much. There certainly is more work to be done to validate this trend, and, if this is so, to work out the point at which quantum models surpass classical machine learning. But there is a possibility that the quantum models can start to compete well with classical models in the not-too-distant future. Already, there are customers that are facing major transitions in their industry, such as the transition to electric cars and autonomous driving in the automotive industry. These companies are willing to invest in new and impactful technologies today that can help them win in the long run — including quantum computing.