Quantum Advantage in Learning from Experiments

post by Dennis Towne (dennis-towne) · 2022-07-27T15:49:25.864Z · LW · GW · 5 comments

This is a link post for https://ai.googleblog.com/2022/06/quantum-advantage-in-learning-from.html

Contents

5 comments

For certain types of problems, quantum computing can provide an exponential speedup for ML algorithms.  This is now being tested on actual QC hardware.

This strikes me as particularly concerning for hard take-off scenarios.  It's a possible way to train dramatically larger AI systems faster, and it's likely to be in use for real problems within the next two decades.

5 comments

Comments sorted by top scores.

comment by Shmi (shminux) · 2022-07-27T18:25:57.613Z · LW(p) · GW(p)

As I understand it, QC provides exponential (or even polynomial) speedup in a very limited subset of interesting problems, each time it exploits the specific structure of the problem at hand, not anything generic. QC will help with physical simulations, of course, but I doubt that is what you mean by "ML algorithms". Do you have a specific ML algorithm in mind where quantum supremacy would make a difference?

comment by gwern · 2022-07-27T16:44:58.568Z · LW(p) · GW(p)

Preprint: https://arxiv.org/abs/2112.00778

Replies from: Charlie Steiner
↑ comment by Charlie Steiner · 2022-07-27T18:53:55.604Z · LW(p) · GW(p)

Ah. The experiment has to return its result as a quantum state.

Yeah this isn't the sort of progress that leads to faster AI algorithms. The thing to watch out for is people trying to combine the speed advantages of quantum annealing with the memory advantages of batched stochastic gradient descent.

My personal expectation is that it won't pan out in time, and we'll mostly just see quantum computers used to simulate quantum systems and break old encryption protocols.

Replies from: shminux
↑ comment by Shmi (shminux) · 2022-07-27T22:28:04.286Z · LW(p) · GW(p)

Yeah, the link is about faster modeling of physical quantum systems, nothing to do with AI/ML as we know it:

The exponential advantage holds in predicting properties of physical systems, performing quantum principal component analysis on noisy states, and learning approximate models of physical dynamics.

comment by Noosphere89 (sharmake-farah) · 2022-07-27T16:52:26.253Z · LW(p) · GW(p)

Yeah, if quantum computers become actually practical to use in AI a hard takeoff is virtually guaranteed and I will be wrong about slow takeoff.

Preregistering a prediction here: If quantum computers are used by AI companies routinely by 2030, hard takeoff will happen by 2040.