The Alliance for Quantum Technologies (AQT), founded by Caltech and AT&T in May 2017 in collaboration with national laboratories and industry partners, is presenting the “INtelligent Quantum NEtworks & Technologies” (INQNET) research program at Supercomputing 2017 in Denver (November 13-16). “The consortium will accelerate progress in quantum science and technologies by bringing together the strengths of government, academia, and industry in a basic science R&D framework,” says Shang-Yi Ch’en Professor of Physics Maria Spiropulu of Caltech.
The INQNET program currently focuses on quantum networks and quantum machine learning. “Quantum networks promise to provide not only ultra-secure communication channels, but could prove critical to developing scalable quantum computing infrastructure,” says, Melissa Arnoldi, president of AT&T Technology Operations.
To this end, the Fermilab quantum network teleportation experiment (FQNET) is being built as part of the INQNET R&D program. “FQNET represents a basic step towards improved entanglement-based networks,” says Cristián Peña, the co-spokesperson of the FQNET experiment and a Lederman Fellow at Fermilab. “With NASA’s Caltech-managed Jet Propulsion Laboratory, we are developing sensors to improve the efficiency and reliability of the entanglement distribution.”
“We are looking forward to testing these technologies in the industry setting,” says Yewon Gim, a senior member of the technical staff of the AT&T Foundry innovation center and an INQNET fellow. “At Fermilab with FQNET we’re trying to push the next phase of instrumentation.”
The INQNET quantum networks group expects the first phase of the FQNET project to produce results by late spring. “The plan is to build eventually a prototype quantum internet, first with nodes in the Chicago area and then around the U.S.,” says Si Xie, a Caltech postdoctoral scholar based at Fermilab who works on the Large Hadron Collider physics and detector upgrades research program.
“We like the model for INQNET,” says Joseph Lykken, the deputy director of Fermliab. “This is a model we know moves very fast. Labs are good for doing things on a larger scale. Industry can bring resources quickly. This can be a very nimble and flexible program.”
“We’re doing a lot of community building,” says Soren Telfer, the director of the Palo Alto AT&T Foundry innovation center. “I already had three interns over the summer who did projects on quantum networks and quantum machine learning.”
“INQNET interns both at graduate and undergraduate level are engaging on quantum technology projects and are rapidly very productive and extremely creative,” says Rishi Pravahan, senior scientist at AT&T, who spearheads the INQNET program at the Foundry.
“With the infusion of expertise from AT&T, Caltech, Fermilab, and others, we are taking a fresh approach in developing quantum networks. We expect that our results will push forward quantum technology while at the same time addressing the greatest fundamental questions in physics,” says Neil Sinclair, an award-winning expert on long-distance quantum technologies who has accepted an offer to join INQNET as a fellow with Harvard and Caltech.
Beyond the near-term demonstrator and benchmarking projects that could have an impact on industry early adoption, the INQNET program also supports a number of long-term projects tackling fundamental challenges in quantum science. According to Lykken and Xie, the INQNET quantum networks group is looking into the possibility of using the FQNET infrastructure to investigate the famous ER=EPR conjecture, which offers a theoretical bridge between the existence of wormholes between a pair of black holes and quantum entanglement.
“Everyone is extremely dedicated and working very hard to accomplish the program’s goals,” says Stacy Kusumolkul, the INQNET program manager at Caltech. “I’m really excited to be part of the starting phase.”
Visit the AQT/INQNET Supercomputing17 booths (#663 and #763) at the Denver convention center Nov 13-Nov 16 2017 and experience the quantum teleportation VR demo developed by NovaVR’s Oliver Lykken.
Written by Mark H. Kim
skusumol AT caltech.edu
Machine learning methods are quickly becoming indispensable in experimental high energy physics, with the Higgs discovery example being a notable use-case where machine learning accelerated the discovery by improving the selection of the Higgs decays in many final states particles, especially electrons and photons which are notoriously plagued with ``fakes'', namely backgrounds that mimic the signal.
The LHC general purpose experiments (ATLAS and CMS) were built to probe the mechanisms of electroweak symmetry breaking and the particle origins of dark matter. Wired up with about a hundred million readout channels each and made up of many thousands of tons of material that interacts with the particles emanating from the LHC’s high-energy proton–proton collisions, the two detectors have already in 2012 managed to capture and reconstruct many rare Higgs boson candidate events. They announced the discovery of the Higgs boson the same year and the discovery resulted in the Nobel prize in Physics being awarded to Higgs and Englert in 2013. The Higgs bosons decay into other particles after about 100 yoctoseconds (10-22) and the collider searches involve several different decay signatures or “channels”. One of them involves the Higgs decaying to two photons.
In this work we apply quantum optimization, for the first time, to a Higgs selection problem, via machine learning implemented on an experimental quantum annealing device with more than 1000 qubits. More specifically, our work shows for the first time how to apply, in practice, quantum annealing for machine learning (QAML). Namely we show how to construct weak classifiers and tie them to physics observables used in the discovery of the Higgs boson. We show that these classifiers are highly resilient against overtraining and MC simulation limitations. In addition the method provides intrinsically a way of picking out the most important variables from a larger set of input variables that are informative in terms of their physics content and interpretation.
Importantly, and also for the first time, we go beyond ground state-based classifiers and we exploit excited states to improve, at negligible algorithmic cost, the performance of our classifier. We focus our studies on performance measures related to classification accuracy measured in terms of receiver operator characteristics. This is a novel perspective that allows us to explore quantum annealers beyond the usual speedup studies and exploit the expected differences in sample distributions between quantum annealers and classical samplers.
In this work we uncover that QAML has a an advantage when learning from small datasets. We plan to explore this further by embedding other high energy physics problems involving real-time decision making and data certification with small datasets, as well as problems dealing with precision measurements where clear interpretability is necessary and systematic errors and correlations are crucial for the discovery of very subtle and rare physics effects. We also envision that quantum and simulated annealing solutions can be used as an initialization and booster to machine learning solutions running on classical machines.
We expect that as more domain science problems are attempted in this architecture and its successors, much more will be learned in terms of the applicability and advantages offered by quantum annealing in hard optimization problems in science and beyond. We are confident that this work will inspire new applications in other domains, whether scientific or commercial, as other quantum annealing, analog quantum computing and other quantum computing testbeds are quickly being developed.
The methods used in this work are generalizable to problems of similar difficulty on this and other quantum annealing architectures.
Q: Why did we pick the specific Higgs optimization problem
The particular Higgs optimization application was chosen because it is simple, with few kinematic variables needed to fully describe the diphoton system. In large measure, this was so that we could test the performance of quantum annealing, as a-priori we did not know what the relative performance would be, and the largest number of variables we could test on the DW2X was only 36. Armed with the results we report here, the community can move on to test moderately larger problems on the newest version of the D-Wave processor, the D-Wave 2000Q (in which 60-variable fully connected problems can be embedded). Using simulated annealing one can also analyze the performance of the algorithm on far larger problems, possibly with dozens of base kinematic variables and thus hundreds or thousands of variables in the Ising problem, which are infeasible to test on any near-term quantum hardware.
Q: What is the goal of this work? usually we hear that these studies are performed to demonstrate “quantum speedup
Given the size of the optimization problem at hand we did not consider speedup as metric for quantumness. Our scope was to formulate a robust, simple, interpretable technique for solving this classification problem which is highly amenable to implementation on quantum annealers, should future generations become large enough, with dense enough connectivity, and with low enough noise rates to achieve advantages over classical solvers.
Since by leaving out a speedup comparison, quantum annealing and simulated annealing would perform identically if both worked perfectly and found the ground state each time, we introduced the idea of constructing classifiers from excited states, knowing that due to sampling differences quantum annealing and simulated annealing would not end up with identical sets of excited states. We wanted to know whether this difference would result in different classification performance. While we cannot claim a statistically significant difference, it is nevertheless noteworthy that at smallest training sizes we did find a slight growing advantage for DW2X over simulated annealing when accounting for an increasing number of excited states, as seen in Fig.5(c) of the paper.
We stress that the technique is easily interpretable and the role of quantum annealing is that of a subroutine for sampling the Ising problem that may in the future have advantages over classical samplers, either when used directly or as a method of seeding classical solvers with-high quality initial states.
Q: Where is the D-Wave machine used in this work
The USC-Lockheed Martin Quantum Computing Center (QCC) based at the USC Information Sciences Institute (ISI) is the longest running quantum computing center in the world. The QCC was launched in November 2011 with a 128-qubit D-Wave One, upgraded in March 2013 with the 512-qubit D-Wave Two, and again in March 2016 with a 1000-qubit D-Wave 2X. The numbers of operational qubits are: 108, 509, 1098, respectively.
The D-Wave 2X Quantum Computing System
We used the current installation at the QCC which is a third-generation D-Wave 2X chip designed to contain up to 1152 superconducting flux qubits, arranged in a two-dimensional 12x12 array of 144 unit cells of 8 qubits each. Within each cell, each qubit is connected to four others, in a K4,4 bipartite graph arrangement. Depending on whether the cell is in a corner or not, the qubits in each cell make 8 or 16 connections to the qubits in neighboring cells. At the time of this work 1098 of the 1152 qubits were functional (95% yield). The D-Wave 2X supports a minimum annealing time of 5µs.
Q: Is the D-Wave a quantum computer?
The D-Wave is not a general purpose digital quantum computer. It is rather an analog quantum computer and specifically a quantum annealer. Both the quantumness and speedup in these devices are intensely scrutinized topics of ongoing research. The D-Wave company implemented the annealer as proprietary scaled-up architecture of flux qubits . Others in the tech industry and academic research are followig-up with research on open quantum computing testbeds and architectures. Beyond the theoretical arguments and demonstrations/papers from the company, we expect the research community in academia and labs to tackle such architectures and produce results on both the instrumentation and the problems encoded and solved in such machines.
Quantum algorithm for solving linear systems of equations, by A. Harrow, A. Hassidim, and S. Lloyd.
Quantum Machine Learning Algorithms: Read the Fine Print, by Scott Aaronson
Why now is the right time to study quantum computing, by Aram Harrow
Written by M. Spiropulu
skusumol AT caltech.edu
"Solving a Higgs optimization problem with quantum annealing for machine learning" is the title of our article that will appear in the October 19 Issue of the Journal Nature.
Form science to business, researchers face a large number of planning and optimization tasks, where they have to choose between different options. By assigning numerical measurements to these options, they can pick the best one that suits their needs. Such a process is called optimization. Decision-making optimization problems are among the hardest and researchers are constantly working to find more efficient ways of solving them, including quantum computing. Annealing -- quantum or simulated-- provides a physics-based technique of arriving to solutions of hard optimization problems.
An example optimization problem in high-energy physics arises in the search for the Higgs boson, which was discovered with the Large Hadron Collider (LHC), the largest particle collider in the world. The LHC is built to enhance our understanding of the fundamental forces of nature, including the mechanism that elementary particles get their mass. The LHC detectors boast about a hundred million "readout channels that collect extremely large amounts of data from the particle collisions. They were employed to chase after the Higgs boson, leading to its discovery in 2012 and to the 2013 Nobel prize in Physics.
The Higgs bosons decay into other particles extremely fast, after about 100 yoctoseconds (10-22 seconds), which made the Higgs search extremely challenging. The researchers seek different decay signatures of the Higgs particle in the data. For example, one signature involves the Higgs boson decaying to two photons. In this work we tackle the optimization problem of maximizing accuracy at classifying events generating two photons in the high energy proton-proton collisions as either true Higgs decays ("our signal") or other, non-Higgs standard model processes ("the background"). Using the physics properties of the photons involved we designed machine-learning classifiers capable of running both on quantum annealers and their simulated counterparts on classical computers.
We have implemented the quantum annealing-based classifiers on the USC 1098 qubit D-Wave machine, a quantum annealer capable of solving certain optimization problems. The D-Wave machine is not a general-purpose quantum computer capable of running common software. Building such a machine (a so-called universal digital quantum computer) is a topic of active research.
We demonstrate in this work that our quantum and simulated annealing classifiers are highly resilient to overtraining (a situation when the classifiers describe the training data tightly but cannot model well the independent test data) and to errors due to inaccurate high-order variable correlations in the relevant Monte Carlo simulations. Additionally the quantum annealing methods pick out the most important variables from a larger set of input variables that are informative in terms of their physics content and interpretation. Our studies reveal a hint that the physics-based classifiers (both quantum and classical annealing) outperform the standard machine learning methods for small training data sets.
This work has already stimulated additional studies applying the new physics-based computation methods to problems in high energy physics and computational biology.
Written by M. Spiropulu
skusumol AT caltech.edu