Korva is a scroll of the internet’s best ideas.
Quantum advantage may need special data
Theoretical results show a proven advantage for certain types of data. The likely candidates are specific distributions: data with particular noise characteristics from the way it was gathered, classes that are spread out in an unusual way, or very rare events that normal machine learning techniques cannot easily detect.
Researchers have been testing practical datasets for years, including biology data from different diseases and cellular mechanisms, to see whether quantum models can realize that advantage in practice. Exactly which kinds of data should show it has not yet been characterized, and neither classical nor quantum computing is standing still, so it is hard to predict where the race ends.
Much of the work already mixes the two. Many quantum neural network architectures use GPUs for the classical layers, put a bit of quantum computing in the middle, and then follow it with more GPU computation.
Why do quantum models still need GPUs?
John Preskill coined the term noisy intermediate-scale quantum, or NISQ, in 2018 for today's machines. They have tens to a few hundred qubits, and errors build up as circuits get deeper. Classical hardware is far faster and more reliable for the heavy, regular arithmetic of neural network layers. So hybrid designs hand quantum processors only the short stretch where they might help.
How do researchers test a claimed advantage?
They compare the quantum method against the best classical method on the same task, not against a weak baseline. Sometimes the comparison erases the gap. In 2018 Ewin Tang, then an undergraduate, published a classical algorithm for a recommendation task that matched the speedup of a celebrated quantum algorithm. This kind of result is called dequantization, and it has been repeated for several other problems.
More from Computerphile
