A Chinese team of physicists have trained a quantum computer to recognise handwritten characters, the first demonstration of “quantum artificial intelligence”
Physicists have long claimed that quantum computers have the potential to dramatically outperform the most powerful conventional processors. The secret sauce at work here is the strange quantum phenomenon of superposition, where a quantum object can exist in two states at the same time.
The advantage comes when one of those states represents a 1 and the other a 0, forming a quantum bit or qubit. In that case, a single quantum object — an atomic nucleus for example— can perform a calculation on two numbers at the same time. Two nuclei can handle 4 numbers, 3 nuclei 8 numbers and 20 nuclei can perform a calculation using more than a million numbers simultaneously. That’s why even a relatively modest quantum computer could dramatically outperform the most advanced supercomputers today.
Quantum physicists have even demonstrated this using small quantum computers crunching a handful of qubits to carry out tasks such as finding the factors of numbers. That’s the kind of calculation that, on a larger scale, could break the codes that currently encrypt top secret communications.
Now physicists have gone a step further. Today, Zhaokai Li and pals at the University of Science and Technology of China in Hefei demonstrate machine learning on a quantum computer for the first time. Their quantum computer can recognise handwritten characters, just as humans can do, in what Li and co are calling the first demonstration of “quantum artificial intelligence”.
Machine learning begins with a set of data to train the device. The idea is to convert each image into a vector by analysing the number of image pixels in each part of the picture.
To keep the experiment simple, the team trained their machine to recognise the difference between a handwritten 6 and a handwritten 9. The vectors representing 6s and 9s can then be compared in this feature space to work out how best to distinguish between them. In effect, the computer finds a hyperplane in the feature space that separates the vectors representing 6s from those representing 9s.
That makes the task of recognising other 6s or 9s straightforward. For each new image of a character, the computer has to decide which side of the dividing line the vector sits.
A key measure of computing performance is the complexity of the problem — the way in which it scales, or takes longer to solve, as the problem gets bigger. The simplest problems are ones which scale in proportion to these variables. So it might take twice as long to process twice as many pictures. Computer scientists say these scale in linear time. Another relatively straightforward type of problem increases with the logarithm of the number of images; it scales in logarithmic time.
But in this case, the problem scales by a factor raised to the power of the number of images and dimensions. In other words, it scales in polynomial time. So as the number of dimensions and images increase, the time it takes to crunch the data increases dramatically.
That’s why quantum computers can help. Last year, a team of quantum theorists devised a quantum algorithm that solves this kind of machine learning problem in logarithmic time rather than polynomial time. That’s a vast speed up. However, their work was entirely theoretical.
It is this algorithm that Li and co have implemented on their quantum computer for the first time.
Their quantum computing machine consists of a small vat of the organic liquid carbon-13-iodotrifluroethylene, a molecule consisting of two carbon atoms attached to three fluorine atoms and one iodine atom. Crucially, one of the carbon atoms is a carbon-13 isotope.
This molecule is handy because each of the three fluorine atoms and the carbon-13 atom can store a single qubit. This works by placing the molecule in a magnetic field to align the spins of the nuclei and then flipping the spins with radio waves. Because each nucleus sits in a slightly different position in the molecule, each can be addressed by slightly different frequencies, a process known as nuclear magnetic resonance.
The spins can also be made to interact with each other so that the molecule acts like a tiny logic gate when zapped by a carefully prepared sequence of radio pulses. In this way the molecule processes data. And because the spins of each nucleus can exist in a superposition of spin up and spin down states, the molecule acts like a tiny quantum computer.
Having processed the quantum information, physicists read out the result by measuring the final states of all the atoms. Because the signal from each molecule is tiny, physicists need an entire vat of them to pick up the processed signal. In this case, an upward peak in the spectrum from the carnon-13 atom indicates the character is a 6 while a downward peak indicates a 9.
The results for the handwritten characters are shown below. “The successful classification shows the ability of our quantum machine to learn and work like an intelligent human,” say Li and co.
That’s an interesting result for artificial intelligence and more broadly for quantum computing. It demonstrates the potential for quantum computation, not just for character recognition, but for other kinds of big data challenges. “This work paves the way to a bright future where the Big Data is processed efficiently in a parallel way provided by quantum mechanics,” say the team.
There are significant challenges ahead, of course. Not least of these is building more powerful quantum computers. The devices that rely on nuclear magnetic resonance cannot handle more than handful of qubits.
So physicists are racing to build quantum computers that can handle significantly more qubits. This is a race with fame and fortune at the end of it. There is no shortage of runners and the team that pulls it off will find an important place in the history of computing and physics in general.
With a few hundred qubits, who knows what quantum artificial intelligence could do.
Ref: arxiv.org/abs/1410.105 4 : Experimental Realization of Quantum Artificial Intelligence
This content was originally published here.