Photonic quantum memristor with a feedback loop beats no-feedback baselines on time-series prediction

[Anchor41]

Quantum Zeitgeist has an article on work that brings a fairly old idea into quantum machine learning. The memristor was proposed in 1971 and only demonstrated in 2008, and it is often held up as a model of a synapse because it retains memory through hysteresis. A team including Mirela Selimovic has now used a photonic quantum memristor as the core of a quantum reservoir computing system, and the write-up describes it as the first implementation of its kind using single photon states

The memristor here is essentially a tunable Mach-Zehnder interferometer. Its internal phase gets updated through a feedback rule based on measurements at one output. That feedback gives the system both nonlinearity and a short-term memory of previous inputs, which is exactly what reservoir computing needs. The quantum reservoir is followed by a simple classical linear regression that does the final training, and that regression adds no extra memory or nonlinearity of its own

They tested it on predicting a smooth nonlinear function and three well-known time series, NARMA, Mackey-Glass and Santa Fe, with vowel recognition done in simulation. According to the researchers, the memristor dynamics improved performance on every task compared with running without the feedback loop. In one monomial prediction test, the quantum memristor model beat classical alternatives while using three free parameters against nine for the classical model

The authors pitch this as a route to more efficient machine learning with fewer resources. They also suggest a memristor like this could serve as the activation layer in quantum neural networks, and even open the door to spiking optical neural networks. They note that equivalent output statistics could be reached with a coherent light source instead of single photons, which is an interesting admission

My reading is that this is a proof of principle rather than a practical machine learning tool. The benchmarks are small and classic, and classical reservoir computers handle them well already. The interesting bit is showing that a single physical node plus feedback can replace the large, randomly connected networks usually needed for reservoir computing

Anyone here working with reservoir computing or neuromorphic hardware? I would like to know whether the parameter advantage is likely to hold up on harder tasks

Quantum Chip Predicts Time Series With A Feedback Loopquantumzeitgeist.com

[DarkEnergy27]

Fewer free parameters for the same or better accuracy is the headline for me. Three versus nine is a tiny example, but efficiency is where machine learning hardware is heading. If that scales, it is a real selling point

[Distant Sienna]

Memristors took nearly four decades to go from theory to hardware, so seeing them in quantum form is a nice twist. The synapse comparison always gets used in press coverage. It is a reasonable analogy as far as analogies go

What matters is whether the feedback trick gives a real advantage over classical photonics. The coherent light comment suggests the quantum side may not be doing all the work here

[AustinTheory]

Mach-Zehnder interferometers with feedback are well understood tools in optics labs. Using one as the whole reservoir is an elegant way to keep the hardware simple

[EdgeRatedR]

Coherent light giving equivalent statistics is a big admission

If a laser can do the job, then the quantum label is doing a lot of marketing work. Still, the architecture could be useful either way

[Ronaldo22]

Mean squared error between 0.92 and 2.76 across tasks doesn't tell me much without the scale of each dataset

Numbers like that need context. Maybe the supplementary material explains it better. I will have a look at the APS paper when I get a moment

[Amy]

Spiking optical networks sound further off than the article implies. Energy efficiency is the real draw, and that is where photonics already has a head start

[Blake_32]

Reservoir computing has always been good at chaotic time series like Mackey-Glass. Classical versions already do very well on those benchmarks. So beating a no-feedback baseline is not the same as beating the best classical approach. I would like to see a head-to-head against a tuned classical reservoir