A Moon-scanning method that can automatically classify important lunar features from telescope images could significantly improve the efficiency of selecting sites for exploration.
There is more than meets the eye to picking a landing or exploration site on the Moon. The visible area of the lunar surface is larger than Russia and is pockmarked by thousands of craters and crisscrossed by canyon-like rilles. The choice of future landing and exploration sites may come down to the most promising prospective locations for construction, minerals or potential energy resources. However, scanning by eye across such a large area, looking for features perhaps a few hundred meters across, is laborious and often inaccurate, which makes it difficult to pick optimal areas for exploration.
Siyuan Chen, Xin Gao and Shuyu Sun, along with colleagues from The Chinese University of Hong Kong, have now applied machine learning and artificial intelligence (AI) to automate the identification of prospective lunar landing and exploration areas.
“We are looking for lunar features like craters and rilles, which are thought to be hotspots for energy resources like uranium and helium-3 — a promising resource for nuclear fusion,” says Chen. “Both have been detected in Moon craters and could be useful resources for replenishing spacecraft fuel.”
Machine learning is a very effective technique for training an AI model to look for certain features on its own. The first problem faced by Chen and his colleagues was that there was no labeled dataset for rilles that could be used to train their model.
“We overcame this challenge by constructing our own training dataset with annotations for both craters and rilles,” says Chen. “To do this, we used an approach called transfer learning to pretrain our rille model on a surface crack dataset with some fine tuning using actual rille masks. Previous approaches require manual annotation for at least part of the input images —our approach does not require human intervention and so allowed us to construct a large high-quality dataset.”
The next challenge was developing a computational approach that could be used to identify both craters and rilles at the same time, something that had not been done before.
“This is a pixel-to-pixel problem for which we need to accurately mask the craters and rilles in a lunar image,” says Chen. “We solved this problem by constructing a deep learning framework called high-resolution-moon-net, which has two independent networks that share the same network architecture to identify craters and rilles simultaneously.”
The team’s approach achieved precision as high as 83.7 percent, higher than existing state-of-the-art methods for crater detection.
Chen, S., Li, Y., Zhang, T., Zhu, X., Sun, S. & Gao, X. Lunar features detection for energy discovery via deep learning. Applied Energy296, 117085 (2021).| article
Galaxy morphology plays an important role in our studies and understanding of galaxy evolution. Structural components such as bulges, disks, spiral arms and bars formed during galaxies aggregated formation histories. As such, morphology is related to other properties that depend on formation and assembly history, such as colour, stellarmass and recent Star Formation Rate (SFR). By looking at the relation between mass and SFR, astronomers have been able to distinguish between three different populations. Most star-forming galaxies belong to the main sequence, and present morphological features typical of spiral or irregular galaxies. Objects in this population are also called Late-Type Galaxies (LTG). We can identify another population with much lower SFR and different shapes, mostly elliptical or bulge-dominated morphologies: we refer to these as Early-Type Galaxies (ETG). The transition between ETG and the main sequence is smoothed by an intermediate and less heavily populated region, called the “green valley”.
The recent increase in the use of machine learning methods has been beneficial for astronomy research, and is of particular interest for extracting information on the evolutionary paths of galaxies from their morphologies. Especially with the exponential rise in the amount of data from modern surveys it has become important to understand and apply intelligent algorithms able to classify galaxies with the same accuracy as human experts, if not even outperforming them.
The algorithms used for image classification typically rely on multiple costly steps, such as the Point Spread Function (PSF) deconvolution and the training and application of complex Convolutional Neural Networks (CNN) of thousands or even millions of parameters. CNNs classify galaxy images by processing different levels of information in each layer, aiming at a progressive recognition of complex features. In this approach, image recognition works well if the objects have clear edges. However galaxies’ outskirts are smooth: even traditional methods like PSF, used to measure structural properties, namely the 2D parametric fitting and the non-parametric analyses, are often prone to inaccuracies due to the difficulty of separating galaxy wings from the background. These boundary effects can be mitigated by using model constraints, but cannot completely prevent inaccurate estimations of structural parameters. Machine learning techniques are also subject to misclassifications for the same reasons, especially with low-resolution images. Another factor to account for when adopting intelligent algorithms is the data management and the speed of the analyses. The increasing volume of available images is difficult to manage and the number of operations processed in CNN models is high. Both training and testing large image data sets requires a lot of time and significant computational costs.
“These limiting factors led us to search for an alternative method, which performs an isophotal analysis of the galaxy light distribution, stores the information in a more manageable data format and performs classification lowering the total computational costs.”
— Tarsitano, first author of the study
Thus, Tarsitano and colleagues, proposed a new approach to extract features from the galaxy images by analysing the elliptical isophotes in their light distribution and collect the information in a sequence. The sequences obtained with this method present definite features allowing a direct distinction between galaxy types, as opposed to smooth Sérsic profiles (Note: Sérsic profile is the most commonly used model. It is a parametric function with parameters describing structural properties such as size, magnitude, ellipticity, inclination and the rate at which light intensity falls off with radius (Sérsic index))..
Then, they trained and classified the sequences with machine learning algorithms, designed through the platform Modulos AutoML, and studied how they optimize the classification task. As a demonstration of this method, they used the second public release of the Dark Energy Survey (DES DR2).
They showed that, by applying it to this sample, they are able to successfully distinguish between early-type and late-type galaxies, for images with signal-to-noise ratio greater then 300. This yields an accuracy of 86% for the early-type galaxies and 93% for the late-type galaxies, which is on par with most contemporary automated image classification approaches.
Finally, they demonstrated that their method allows for galaxy images to be accurately classified and is faster than other approaches. Data dimensionality reduction also implies a significant lowering in computational cost. While, in the perspective of future data sets obtained with e.g. Euclid and the Vera Rubin Observatory (VRO), their work represents a path towards using a well-tested and widely used platform from industry in efficiently tackling galaxy classification problems at the peta-byte scale.
“In the future, we will expand upon our promising results by developing a more robust isophotal measurement approach to focus on performance at low S/N, and target higher context features, such as bars, spiral arms and clumps.”
Reference: F. Tarsitano, C. Bruderer, K. Schawinski, W. G. Hartley, “Image feature extraction and galaxy classification: a novel and efficient approach with automated machine learning”, Arxiv, pp. 1-9, 2021. https://arxiv.org/abs/2105.01070
Note for editors of other websites: To reuse this article fully or partially, kindly give credit either to our author S. Aman or provide link of our article
Using neural networks, researchers can now simulate universes in a fraction of the time, advancing the future of physics research
A universe evolves over billions upon billions of years, but researchers have developed a way to create a complex simulated universe in less than a day. The technique, published in this week’s Proceedings of the National Academy of Sciences, brings together machine learning, high-performance computing and astrophysics and will help to usher in a new era of high-resolution cosmology simulations.
Cosmological simulations are an essential part of teasing out the many mysteries of the universe, including those of dark matter and dark energy. But until now, researchers faced the common conundrum of not being able to have it all — simulations could focus on a small area at high resolution, or they could encompass a large volume of the universe at low resolution.
Carnegie Mellon University Physics Professors Tiziana Di Matteo and Rupert Croft, Flatiron Institute Research Fellow Yin Li, Carnegie Mellon Ph.D. candidate Yueying Ni, University of California Riverside Professor of Physics and Astronomy Simeon Bird and University of California Berkeley’s Yu Feng surmounted this problem by teaching a machine learning algorithm based on neural networks to upgrade a simulation from low resolution to super resolution.
“Cosmological simulations need to cover a large volume for cosmological studies, while also requiring high resolution to resolve the small-scale galaxy formation physics, which would incur daunting computational challenges. Our technique can be used as a powerful and promising tool to match those two requirements simultaneously by modeling the small-scale galaxy formation physics in large cosmological volumes,” said Ni, who performed the training of the model, built the pipeline for testing and validation, analyzed the data and made the visualization from the data.
The trained code can take full-scale, low-resolution models and generate super-resolution simulations that contain up to 512 times as many particles. For a region in the universe roughly 500 million light-years across containing 134 million particles, existing methods would require 560 hours to churn out a high-resolution simulation using a single processing core. With the new approach, the researchers need only 36 minutes.
The results were even more dramatic when more particles were added to the simulation. For a universe 1,000 times as large with 134 billion particles, the researchers’ new method took 16 hours on a single graphics processing unit. Using current methods, a simulation of this size and resolution would take a dedicated supercomputer months to complete.
Reducing the time it takes to run cosmological simulations “holds the potential of providing major advances in numerical cosmology and astrophysics,” said Di Matteo. “Cosmological simulations follow the history and fate of the universe, all the way to the formation of all galaxies and their black holes.”
Scientists use cosmological simulations to predict how the universe would look in various scenarios, such as if the dark energy pulling the universe apart varied over time. Telescope observations then confirm whether the simulations’ predictions match reality.
“With our previous simulations, we showed that we could simulate the universe to discover new and interesting physics, but only at small or low-res scales,” said Croft. “By incorporating machine learning, the technology is able to catch up with our ideas.”
“The universe is the biggest data sets there is — artificial intelligence is the key to understanding the universe and revealing new physics,” said Scott Dodelson, professor and head of the department of physics at Carnegie Mellon University and director of the NSF Planning Institute. “This research illustrates how the NSF Planning Institute for Artificial Intelligence will advance physics through artificial intelligence, machine learning, statistics and data science.”
“It’s clear that AI is having a big effect on many areas of science, including physics and astronomy,” said James Shank, a program director in NSF’s Division of Physics. “Our AI planning Institute program is working to push AI to accelerate discovery. This new result is a good example of how AI is transforming cosmology.”
To create their new method, Ni and Li harnessed these fields to create a code that uses neural networks to predict how gravity moves dark matter around over time. The networks take training data, run calculations and compare the results to the expected outcome. With further training, the networks adapt and become more accurate.
The specific approach used by the researchers, called a generative adversarial network, pits two neural networks against each other. One network takes low-resolution simulations of the universe and uses them to generate high-resolution models. The other network tries to tell those simulations apart from ones made by conventional methods. Over time, both neural networks get better and better until, ultimately, the simulation generator wins out and creates fast simulations that look just like the slow conventional ones.
“We couldn’t get it to work for two years,” Li said, “and suddenly it started working. We got beautiful results that matched what we expected. We even did some blind tests ourselves, and most of us couldn’t tell which one was ‘real’ and which one was ‘fake.’”
Despite only being trained using small areas of space, the neural networks accurately replicated the large-scale structures that only appear in enormous simulations.
The simulations didn’t capture everything, though. Because they focused on dark matter and gravity, smaller-scale phenomena — such as star formation, supernovae and the effects of black holes — were left out. The researchers plan to extend their methods to include the forces responsible for such phenomena, and to run their neural networks ‘on the fly’ alongside conventional simulations to improve accuracy.
The research was powered by the Frontera supercomputer at the Texas Advanced Computing Center (TACC), the fastest academic supercomputer in the world. The team is one of the largest users of this massive computing resource, which is funded by the NSF Office of Advanced Cyberinfrastructure.
This research was funded by the NSF, the NSF AI Institute: Physics of the Future and NASA.
Featured image: The leftmost simulation ran at low resolution. Using machine learning, researchers upscaled the low-res model to create a high-resolution simulation (right). That simulation captures the same details as a conventional high-res model (middle) while requiring significantly fewer computational resources. Credit: Y. Li et al./Proceedings of the National Academy of Sciences 2021
Reference: Yin Li, Yueying Ni, Rupert A. C. Croft, Tiziana Di Matteo, Simeon Bird, Yu Feng, “AI-assisted superresolution cosmological simulations”, Proceedings of the National Academy of Sciences May 2021, 118 (19) e2022038118; DOI: https://doi.org/10.1073/pnas.2022038118
Protocol to reverse engineer Hamiltonian models advances automation of quantum devices
Scientists from the University of Bristol’s Quantum Engineering Technology Labs (QETLabs) have developed an algorithm that provides valuable insights into the physics underlying quantum systems – paving the way for significant advances in quantum computation and sensing, and potentially turning a new page in scientific investigation.
In physics, systems of particles and their evolution are described by mathematical models, requiring the successful interplay of theoretical arguments and experimental verification. Even more complex is the description of systems of particles interacting with each other at the quantum mechanical level, which is often done using a Hamiltonian model. The process of formulating Hamiltonian models from observations is made even harder by the nature of quantum states, which collapse when attempts are made to inspect them.
In the paper, Learning models of quantum systems from experiments, published in Nature Physics, quantum mechanics from Bristol’s QET Labs describe an algorithm which overcomes these challenges by acting as an autonomous agent, using machine learning to reverse engineer Hamiltonian models.
The team developed a new protocol to formulate and validate approximate models for quantum systems of interest. Their algorithm works autonomously, designing and performing experiments on the targeted quantum system, with the resultant data being fed back into the algorithm. It proposes candidate Hamiltonian models to describe the target system, and distinguishes between them using statistical metrics, namely Bayes factors.
Excitingly, the team were able to successfully demonstrate the algorithm’s ability on a real-life quantum experiment involving defect centres in a diamond, a well-studied platform for quantum information processing and quantum sensing.
The algorithm could be used to aid automated characterisation of new devices, such as quantum sensors. This development therefore represents a significant breakthrough in the development of quantum technologies.
“Combining the power of today’s supercomputers with machine learning, we were able to automatically discover structure in quantum systems. As new quantum computers/simulators become available, the algorithm becomes more exciting: first it can help to verify the performance of the device itself, then exploit those devices to understand ever-larger systems,” said Brian Flynn from the University of Bristol’s QETLabs and Quantum Engineering Centre for Doctoral Training.
“This level of automation makes it possible to entertain myriads of hypothetical models before selecting an optimal one, a task that would be otherwise daunting for systems whose complexity is ever increasing,” said Andreas Gentile, formerly of Bristol’s QETLabs, now at Qu & Co.
“Understanding the underlying physics and the models describing quantum systems, help us to advance our knowledge of technologies suitable for quantum computation and quantum sensing,” said Sebastian Knauer, also formerly of Bristol’s QETLabs and now based at the University of Vienna’s Faculty of Physics.
Anthony Laing, co-Director of QETLabs and Associate Professor in Bristol’s School of Physics, and an author on the paper, praised the team: “In the past we have relied on the genius and hard work of scientists to uncover new physics. Here the team have potentially turned a new page in scientific investigation by bestowing machines with the capability to learn from experiments and discover new physics. The consequences could be far reaching indeed.”
The next step for the research is to extend the algorithm to explore larger systems, and different classes of quantum models which represent different physical regimes or underlying structures.
Paper: Gentile, A.A., Flynn, B., Knauer, S. et al. Learning models of quantum systems from experiments. Nat. Phys. (2021). Link when paper is live: https://dx.doi.org/10.1038/s41567-021-01201-7
Using data from the Dark Energy Survey, researchers from the Department of Physics & Astronomy produced the largest catalog of galaxy morphology classifications to date.
Research from Penn’s Department of Physics and Astronomy has produced the largest catalog of galaxy morphology classification to date. Led by former postdocs Jesús Vega-Ferrero and Helena Domínguez Sánchez, who worked with professor Mariangela Bernardi, this catalog of 27 million galaxy morphologies provides key insights into the evolution of the universe. The study was published in Monthly Notices of the Royal Astronomical Society.
The researchers used data from the Dark Energy Survey (DES), an international research program whose goal is to image one-eighth of the sky to better understand dark energy’s role in the accelerating expansion of the universe.
A byproduct of this survey is that the DES data contains many more images of distant galaxies than other surveys to date. “The DES images show us what galaxies looked like more than 6 billion years ago,” says Bernardi.
And because DES has millions of high-quality images of astronomical objects, it’s the perfect dataset for studying galaxy morphology. “Galaxy morphology is one of the key aspects of galaxy evolution. The shape and structure of galaxies has a lot of information about the way they were formed, and knowing their morphologies gives us clues as to the likely pathways for the formation of the galaxies,” Domínguez Sánchez says.
Previously, the researchers had published a morphological catalog for more than 600,000 galaxies from the Sloan Digital Sky Survey (SDSS). To do this, they developed a convolutional neural network, a type of machine learning algorithm, that was able to automatically categorize whether a galaxy belonged to one of two major groups: spiral galaxies, which have a rotating disk where new stars are born, and elliptical galaxies, which are larger, and made of older stars which move more randomly than their spiral counterparts.
But the catalog developed using the SDSS dataset was primarily made of bright, nearby galaxies, says Vega-Ferrero. In their latest study, the researchers wanted to refine their neural network model to be able to classify fainter, more distant galaxies. “We wanted to push the limits of morphological classification and trying to go beyond, to fainter objects or objects that are farther away,” Vega-Ferrero says.
To do this, the researchers first had to train their neural network model to be able to classify the more pixelated images from the DES dataset. They first created a training model with previously known morphological classifications, comprised of a set of 20,000 galaxies that overlapped between DES and SDSS. Then, they created simulated versions of new galaxies, mimicking what the images would look like if they were farther away using code developed by staff scientist Mike Jarvis.
Images of a simulated spiral (top) and elliptical galaxy at varying image quality and redshift levels, illustrating how fainter and more distant galaxies might look within the DES dataset. (Image: Jesus Vega-Ferrero and Helena Dominguez-Sanchez).
Once the model was trained and validated on both simulated and real galaxies, it was applied to the DES dataset, and the resulting catalog of 27 million galaxies includes information on the probability of an individual galaxy being elliptical or spiral. The researchers also found that their neural network was 97% accurate at classifying galaxy morphology, even for galaxies that were too faint to classify by eye.
“We pushed the limits by three orders of magnitude, to objects that are 1,000 times fainter than the original ones,” Vega-Ferrero says. “That is why we were able to include so many more galaxies in the catalog.”
“Catalogs like this are important for studying galaxy formation,” Bernardi says about the significance of this latest publication. “This catalog will also be useful to see if the morphology and stellar populations tell similar stories about how galaxies formed.”
For the latter point, Domínguez Sánchez is currently combining their morphological estimates with measures of the chemical composition, age, star-formation rate, mass, and distance of the same galaxies. Incorporating this information will allow the researchers to better study the relationship between galaxy morphology and star formation, work that will be crucial for a deeper understanding of galaxy evolution.
Bernardi says that there are a number of open questions about galaxy evolution that both this new catalog, and the methods developed to create it, can help address. The upcoming LSST/Rubin survey, for example, will use similar photometry methods to DES but will have the capability of imaging even more distant objects, providing an opportunity to gain even deeper understanding of the evolution of the universe.
This research was supported by NSF Grant AST-1816330.
Featured image: An image of NGC 1365 collected by the Dark Energy Survey. Also known as the Great Barred Spiral Galaxy, NGC 1365 is an example of a spiral galaxy and is located about 56 million light-years away. (Image: DECam, DES Collaboration)
Reference: J Vega-Ferrero, H Domínguez Sánchez, M Bernardi, M Huertas-Company, R Morgan, B Margalef, M Aguena, S Allam, J Annis, S Avila, D Bacon, E Bertin, D Brooks, A Carnero Rosell, M Carrasco Kind, J Carretero, A Choi, C Conselice, M Costanzi, L N da Costa, M E S Pereira, J De Vicente, S Desai, I Ferrero, P Fosalba, J Frieman, J García-Bellido, D Gruen, R A Gruendl, J Gschwend, G Gutierrez, W G Hartley, S R Hinton, D L Hollowood, K Honscheid, B Hoyle, M Jarvis, A G Kim, K Kuehn, N Kuropatkin, M Lima, M A G Maia, F Menanteau, R Miquel, R L C Ogando, A Palmese, F Paz-Chinchón, A A Plazas, A K Romer, E Sanchez, V Scarpine, M Schubnell, S Serrano, I Sevilla-Noarbe, M Smith, E Suchyta, M E C Swanson, G Tarle, F Tarsitano, C To, D L Tucker, T N Varga, R D Wilkinson, Pushing automated morphological classifications to their limits with the Dark Energy Survey, Monthly Notices of the Royal Astronomical Society, 2021;, stab594, https://doi.org/10.1093/mnras/stab594
Army-funded researchers demonstrated a machine learning approach that corrects quantum information in systems composed of photons, improving the outlook for deploying quantum sensing and quantum communications technologies on the battlefield.
When photons are used as the carriers of quantum information to transmit data, that information is often distorted due to environment fluctuations destroying the fragile quantum states necessary to preserve it.
Researchers from Louisiana State University exploited a type of machine learning to correct for information distortion in quantum systems composed of photons. Published in Advanced Quantum Technologies, the team demonstrated that machine learning techniques using the self-learning and self-evolving features of artificial neural networks can help correct distorted information. This results outperformed traditional protocols that rely on conventional adaptive optics.
“We are still in the fairly early stages of understanding the potential for machine learning techniques to play a role in quantum information science,” said Dr. Sara Gamble, program manager at the Army Research Office, an element of U.S. Army Combat Capabilities Development Command, known as DEVCOM, Army Research Laboratory. “The team’s result is an exciting step forward in developing this understanding, and it has the potential to ultimately enhance the Army’s sensing and communication capabilities on the battlefield.”
For this research, the team used a type of neural network to correct for distorted spatial modes of light at the single-photon level.
“The random phase distortion is one of the biggest challenges in using spatial modes of light in a wide variety of quantum technologies, such as quantum communication, quantum cryptography, and quantum sensing,” said Narayan Bhusal, doctoral candidate at LSU. “Our method is remarkably effective and time-efficient compared to conventional techniques. This is an exciting development for the future of free-space quantum technologies.”
According to the research team, this smart quantum technology demonstrates the possibility of encoding of multiple bits of information in a single photon in realistic communication protocols affected by atmospheric turbulence.
“Our technique has enormous implications for optical communication and quantum cryptography,” said Omar Magaña‐Loaiza, assistant professor of physics at LSU. “We are now exploring paths to implement our machine learning scheme in the Louisiana Optical Network Initiative to make it smart, more secure, and quantum.”
Featured image: Army-funded researchers demonstrate a machine learning approach that corrects quantum information distortion in systems composed of photons, improving the outlook for deploying quantum sensing and quantum communications technologies on the battlefield. (Courtesy: LSU)
Reference: Bhusal, N., Lohani, S., You, C., Hong, M., Fabre, J., Zhao, P., Knutson, E.M., Glasser, R.T. and Magaña‐Loaiza, O.S. (2021), Front Cover: Spatial Mode Correction of Single Photons Using Machine Learning (Adv. Quantum Technol. 3/2021). Adv. Quantum Technol., 4: 2170031. https://doi.org/10.1002/qute.202170031
New theorems put quantum machine learning on rigorous footing and identify the key issue that will determine whether it will provide quantum speedup
The work solves a key problem of useability for quantum machine learning. We rigorously proved the conditions under which certain architectures of variational quantum algorithms will or will not have barren plateaus as they are scaled up.- Marco Cerezo
Many machine learning algorithms on quantum computers suffer from the dreaded “barren plateau” of unsolvability, where they run into dead ends on optimization problems. This challenge had been relatively unstudied—until now. Rigorous theoretical work has established theorems that guarantee whether a given machine learning algorithm will work as it scales up on larger computers.
“The work solves a key problem of useability for quantum machine learning. We rigorously proved the conditions under which certain architectures of variational quantum algorithms will or will not have barren plateaus as they are scaled up,” said Marco Cerezo, lead author on the paper published in Nature Communications today by a Los Alamos National Laboratory team. Cerezo is a post doc researching quantum information theory at Los Alamos. “With our theorems, you can guarantee that the architecture will be scalable to quantum computers with a large number of qubits.”
“Usually the approach has been to run an optimization and see if it works, and that was leading to fatigue among researchers in the field,” said Patrick Coles, a coauthor of the study. Establishing mathematical theorems and deriving first principles takes the guesswork out of developing algorithms.
The Los Alamos team used the common hybrid approach for variational quantum algorithms, training and optimizing the parameters on a classical computer and evaluating the algorithm’s cost function, or the measure of the algorithm’s success, on a quantum computer.
Machine learning algorithms translate an optimization task—say, finding the shortest route for a traveling salesperson through several cities—into a cost function, said coauthor Lukasz Cincio. That’s a mathematical description of a function that will be minimized. The function reaches its minimum value only if you solve the problem.
Most quantum variational algorithms initiate their search randomly and evaluate the cost function globally across every qubit, which often leads to a barren plateau.
“We were able to prove that, if you choose a cost function that looks locally at each individual qubit, then we guarantee that the scaling won’t result in an impossibly steep curve of time versus system size, and therefore can be trained,” Coles said.
A quantum variational algorithm sets up a problem-solving landscape where the peaks represent the high energy points of the system, or problem, and the valleys are the low energy values. The answer lies in the deepest valley. That’s the ground state, represented by the minimized cost function. To find the solution, the algorithm trains itself about the landscape, thereby navigating to the low spot.
“People have been proposing quantum neural networks and benchmarking them by doing small-scale simulations of 10s (or fewer) few qubits,” Cerezo said. “The trouble is, you won’t see the barren plateau with a small number of qubits, but when you try to scale up to more qubits, it appears. Then the algorithm has to be reworked for a larger quantum computer.”
A barren plateau is a trainability problem that occurs in machine learning optimization algorithms when the problem-solving space turns flat as the algorithm is run. In that situation, the algorithm can’t find the downward slope in what appears to be a featureless landscape and there’s no clear path to the energy minimum. Lacking landscape features, the machine learning can’t train itself to find the solution.
“If you have a barren plateau, all hope of quantum speedup or quantum advantage is lost,” Cerezo said.
The Los Alamos team’s breakthrough takes an important step toward quantum advantage, when a quantum computer performs a task that would take infinitely long on a classical computer. Achieving quantum advantage hinges in the short term on scaling up variational quantum algorithms. These algorithms have the potential so solve practical problems when quantum computers of 100 qubits or more become available—hopefully soon. Quantum computers currently max out at 65 qubits. A qubit is the basic unit of information in a quantum computer, as bits are in a classical digital computer.
“The hottest topic in noisy intermediate-scale quantum computers is variational quantum algorithms, or quantum machine learning and quantum neural networks,” Coles said. “They have been proposed for applications from solving the structure of a molecule in chemistry to simulating the dynamics of atoms and molecules and factoring numbers.”
The paper: “Cost function dependent barren plateaus in shallow parametrized quantum circuits,” Nature Communications, by M. Cerezo, Akira Sone, Tyler Volkoff, Lukasz Cincio, and Patrick J. Coles. DOI: 10.1038/s41467-021-21728-w
Funding: Los Alamos National Laboratory’s Laboratory Directed Research and Development (LDRD), the Center for Nonlinear Studies at Los Alamos, the ASC Beyond Moore’s Law project at Los Alamos, and the U.S. Department of Energy (DOE), Office of Science, Office of Advanced Scientific Computing Research.
A Florida State University professor’s research could help quantum computing fulfill its promise as a powerful computational tool.
William Oates, the Cummins Inc. Professor in Mechanical Engineering and chair of the Department of Mechanical Engineering at the FAMU-FSU College of Engineering, and postdoctoral researcher Guanglei Xu found a way to automatically infer parameters used in an important quantum Boltzmann machine algorithm for machine learning applications.
The work could help build artificial neural networks that could be used for training computers to solve complicated, interconnected problems like image recognition, drug discovery and the creation of new materials.
“There’s a belief that quantum computing, as it comes online and grows in computational power, can provide you with some new tools, but figuring out how to program it and how to apply it in certain applications is a big question,” Oates said.
Quantum bits, unlike binary bits in a standard computer, can exist in more than one state at a time, a concept known as superposition. Measuring the state of a quantum bit — or qubit — causes it to lose that special state, so quantum computers work by calculating the probability of a qubit’s state before it is observed.
Specialized quantum computers known as quantum annealers are one tool for doing this type of computing. They work by representing each state of a qubit as an energy level. The lowest energy state among its qubits gives the solution to a problem. The result is a machine that could handle complicated, interconnected systems that would take a regular computer a very long time to calculate — like building a neural network.
One way to build neural networks is by using a restricted Boltzmann machine, an algorithm that uses probability to learn based on inputs given to the network. Oates and Xu found a way to automatically calculate an important parameter associated with effective temperature that is used in that algorithm. Restricted Boltzmann machines typically guess at that parameter instead, which requires testing to confirm and can change whenever the computer is asked to investigate a new problem.
“That parameter in the model replicates what the quantum annealer is doing,” Oates said. “If you can accurately estimate it, you can train your neural network more effectively and use it for predicting things.”
This research was supported by Cummins Inc. and used resources of the Oak Ridge Leadership Computing Facility, which is a DOE Office of Science User Facility.
Featured image: William Oates, the Cummins Inc. Professor in Mechanical Engineering and chair of the Department of Mechanical Engineering at the FAMU-FSU College of Engineering. (FAMU-FSU College of Engineering/Mark Wallheisier)
A machine learning-based tool was able to predict the risk of malignancy among patients presenting with multiple pulmonary nodules and outperformed human experts, previously validated mathematical models, and a previously established artificial intelligence tool, according to results published in Clinical Cancer Research, a journal of the American Association for Cancer Research.
Tools currently available can predict malignancy in patients with single nodules; predictive tools for patients presenting with multiple nodules are limited.
“With the adoption of widespread use of thoracic computed tomography (CT) for lung cancer screening, the detection of multiple pulmonary nodules has become increasingly common,” said study author Kezhong Chen, MD, vice professor in the Department of Thoracic Surgery at Peking University People’s Hospital in China. Among patients presenting with a pulmonary nodule on a CT scan in a previous lung cancer screening trial, roughly 50 percent presented with multiple nodules, Chen said.
“Current guidelines recommend the use of clinical models that incorporate nodule and sociodemographic features to estimate the probability of cancer prior to surgical treatment, and while there are several tools for patients that present with a single nodule, no such tool currently exists for patients with multiple nodules, representing an urgent medical need,” Chen added.
To address this unmet need, the researchers set out to develop a machine learning-based model to predict the probability of lung malignancy among patients presenting with multiple pulmonary nodules. First, the study authors used data from a training cohort of 520 patients (comprising 1,739 nodules) who were treated at Peking University People’s Hospital between January 2007 and December 2018. Using both radiographical nodule characteristics and sociodemographic variables, the authors developed a model, termed PKU-M, to predict the probability of cancer. The performance of the model was evaluated by calculating the area under the curve (AUC), where a score of 1 corresponds to a perfect prediction. In the training cohort, the model achieved an AUC of 0.91. Some of the top predictive features of the model included nodule size, nodule count, nodule distribution, and age of the patient.
The model was then validated using data from a cohort of 220 patients (comprising 583 nodules) who underwent surgical treatment in six independent hospitals in China and Korea between January 2016 and December 2018. The performance of the PKU-M model in this cohort was similar to its performance in the training cohort, with an AUC of 0.89. The researchers also compared the performance of their model with four prior logistic regression-based models that were developed for the prediction of lung cancer. The PKU-M model outperformed all four of the prior models, whose AUC values ranged from 0.68 to 0.81.
Finally, the researchers conducted a prospective comparison between the PKU-M model, three thoracic surgeons, a radiologist, and a previously established artificial intelligence tool for the diagnosis of lung cancer called RX. This comparison was conducted on a separate cohort of 78 patients (comprising 200 nodules) who underwent surgical treatment at four independent hospitals in China between January 2019 and March 2019. Similar to the training and validation cohorts, the performance of the PKU-M model achieved an AUC of 0.87, which was higher than that from the surgeons (with AUCs ranging from 0.73 to 0.79), the radiologist (AUC of 0.75), and the RX model (AUC of 0.76).
“The increasing detection rate of multiple pulmonary nodules has led to an emerging problem for lung cancer diagnosis,” said study author Young Tae Kim, MD, PhD, professor in the Department of Thoracic and Cardiovascular Surgery at Seoul National University Hospital and the Seoul National University College of Medicine in the Republic of Korea. “Because many nodules are found to be benign either after long-term follow-up or surgery, it is important to carefully evaluate these nodules prior to invasive procedures. Our prediction model, which was exclusively established for patients with multiple nodules, can help not only mitigate unnecessary surgery but also facilitate the diagnosis and treatment of lung cancer.”
“Models are developed to assist in clinical diagnosis, which means that they should be practical,” said study author Jun Wang, MD, professor in the Department of Thoracic Surgery at Peking University People’s Hospital. “We therefore designed a web-based version of the PKU-M model, where clinicians can input several clinical and radiological characteristics and the software will automatically calculate the risk of malignancy in a specific patient. This tool can quickly generate an objective diagnosis and can aid in clinical decision-making.”
Because this study only used data from Asian patients, it may not be generalizable to a Western population or other populations, representing a limitation of this study.
This study was funded by the National Natural Science Foundation of China and Peking University People’s Hospital.
Reference: Kezhong Chen, Yuntao Nie, Samina Park, Kai Zhang, Yangming Zhang, Yuan Liu, Bengang Hui, Lixin Zhou, Xun Wang, Qingyi Qi, Hao Li, Guannan Kang, Yuqing Huang, Yingtai Chen, Jiabao Liu, Jian Cui, Mingru Li, In Kyu Park, Chang Hyun Kang, Haifeng Shen, Yingshun Yang, Tian Guan, Yaxiao Zhang, Fan Yang, Young Tae Kim and Jun Wang, “Development and Validation of Machine Learning–based Model for the Prediction of Malignancy in Multiple Pulmonary Nodules: Analysis from Multicentric Cohorts”, Clin Cancer Res 2021. DOI: 10.1158/1078-0432.CCR-20-4007
UR 