Connect with us

Horoscopo

Las transformadas de Fourier revelan cómo la IA aprende física compleja

Published

on

Un nuevo estudio ha encontrado que el análisis de Fourier, una técnica matemática que existe desde hace 200 años, se puede utilizar para revelar información importante sobre cómo las redes neuronales profundas aprenden a realizar tareas físicas complejas, como modelar el clima y la turbulencia. Esta investigación destaca el potencial del análisis de Fourier como herramienta para comprender mejor el funcionamiento interno de la inteligencia artificial y podría tener implicaciones importantes para el desarrollo de algoritmos de aprendizaje automático más eficientes.

La ‘caja negra’ de la IA científica no es rival para el método de 200 años

Las transformadas de Fourier revelan qué tan bien la red neuronal profunda aprende física compleja.

Una de las herramientas más antiguas de la física computacional, una técnica matemática de 200 años conocida como análisis de Fourier – puede revelar información crucial sobre cómo una forma de inteligencia artificial llamada red neuronal profunda aprende a realizar tareas que involucran física compleja como modelar el clima y la turbulencia, según un nuevo estudio.

El descubrimiento realizado por investigadores de ingeniería mecánica de la Universidad de Rice se describe en un estudio de acceso abierto publicado en la revista Nexo PNASuna publicación hermana de procedimientos de la Academia Nacional de Ciencias.

«Este es el primer marco riguroso para explicar y guiar el uso de redes neuronales profundas para sistemas dinámicos complejos como el clima», dijo el autor correspondiente del estudio, Pedram Hassanzadeh. «Esto podría acelerar drásticamente el uso del aprendizaje científico profundo en la ciencia del clima y conducir a proyecciones mucho más confiables del cambio climático».

La IA predice cómo cambiarán los flujos con el tiempo

Los investigadores de la Universidad de Rice entrenaron una forma de inteligencia artificial llamada red neuronal de aprendizaje profundo para reconocer flujos complejos de aire o agua y predecir cómo cambiarán los flujos con el tiempo. Esta imagen ilustra las diferencias sustanciales en la escala de características que muestra el modelo durante el entrenamiento (arriba) y las características que aprende a reconocer (abajo) para hacer sus predicciones. Crédito: Imagen cortesía de P. Hassanzadeh/Universidad Rice

En el artículo, Hassanzadeh, Adam Subel y Ashesh Chattopadhyay, ambos exalumnos, y Yifei Guan, investigador asociado postdoctoral, detallaron su uso del análisis de Fourier para estudiar una red neuronal de aprendizaje profundo que ha sido entrenada para reconocer los complejos flujos de aire en el atmósfera. o agua en el océano y predecir cómo estos flujos cambiarían con el tiempo. Su análisis reveló «no solo lo que había aprendido la red neuronal, sino que también nos permitió conectar directamente lo que había aprendido la red con la física del sistema complejo que estaba modelando», dijo Hassanzadeh.

“Las redes neuronales profundas son tristemente dificil de entender ya menudo son vistos como ‘cajas negras’”, dijo. “Esta es una de las principales preocupaciones con el uso de redes neuronales profundas en aplicaciones científicas. La otra es la generalización: estas redes no pueden funcionar para un sistema diferente de aquel para el que fueron entrenadas. . »

Espectros de Fourier de la mayoría de los núcleos modificados de DNN reciclado

El entrenamiento de redes neuronales profundas de última generación requiere una gran cantidad de datos, y la carga de reentrenamiento, con los métodos actuales, sigue siendo significativa. Después de entrenar y volver a entrenar una red de aprendizaje profundo para realizar diferentes tareas que involucran física compleja, los investigadores de la Universidad de Rice utilizaron el análisis de Fourier para comparar los 40,000 núcleos de las dos iteraciones y encontraron que más del 99% eran similares. Esta ilustración muestra los espectros de Fourier de los cuatro núcleos que diferían más antes (izquierda) y después (derecha) del reentrenamiento. Los resultados demuestran el potencial del método para identificar vías de reciclaje más eficientes que requieren muchos menos datos. Crédito: Imagen cortesía de P. Hassanzadeh/Universidad Rice

Hassanzadeh dijo que el marco analítico que presenta su equipo en el documento «abre la caja negra, nos permite mirar dentro para comprender qué han aprendido las redes y por qué, y también nos permite relacionar esto con la física del sistema que se ha aprendido». .

Subel, el autor principal del estudio, comenzó la investigación como estudiante universitario en Rice y ahora es estudiante de posgrado en[{» attribute=»»>New York University. He said the framework could be used in combination with techniques for transfer learning to “enable generalization and ultimately increase the trustworthiness of scientific deep learning.”

While many prior studies had attempted to reveal how deep learning networks learn to make predictions, Hassanzadeh said he, Subel, Guan and Chattopadhyay chose to approach the problem from a different perspective.

Pedram Hassanzadeh

Pedram Hassanzadeh. Credit: Rice Universit

“The common machine learning tools for understanding neural networks have not shown much success for natural and engineering system applications, at least such that the findings could be connected to the physics,” Hassanzadeh said. “Our thought was, ‘Let’s do something different. Let’s use a tool that’s common for studying physics and apply it to the study of a neural network that has learned to do physics.”

He said Fourier analysis, which was first proposed in the 1820s, is a favorite technique of physicists and mathematicians for identifying frequency patterns in space and time.

“People who do physics almost always look at data in the Fourier space,” he said. “It makes physics and math easier.”

For example, if someone had a minute-by-minute record of outdoor temperature readings for a one-year period, the information would be a string of 525,600 numbers, a type of data set physicists call a time series. To analyze the time series in Fourier space, a researcher would use trigonometry to transform each number in the series, creating another set of 525,600 numbers that would contain information from the original set but look quite different.

“Instead of seeing temperature at every minute, you would see just a few spikes,” Subel said. “One would be the cosine of 24 hours, which would be the day and night cycle of highs and lows. That signal was there all along in the time series, but Fourier analysis allows you to easily see those types of signals in both time and space.”

Based on this method, scientists have developed other tools for time-frequency analysis. For example, low-pass transformations filter out background noise, and high-pass filters do the inverse, allowing one to focus on the background.

Adam Subel

Adam Subel. Credit: Rice University

Hassanzadeh’s team first performed the Fourier transformation on the equation of its fully trained deep-learning model. Each of the model’s approximately 1 million parameters act like multipliers, applying more or less weight to specific operations in the equation during model calculations. In an untrained model, parameters have random values. These are adjusted and honed during training as the algorithm gradually learns to arrive at predictions that are closer and closer to the known outcomes in training cases. Structurally, the model parameters are grouped in some 40,000 five-by-five matrices, or kernels.

“When we took the Fourier transform of the equation, that told us we should look at the Fourier transform of these matrices,” Hassanzadeh said. “We didn’t know that. Nobody has done this part ever before, looked at the Fourier transforms of these matrices and tried to connect them to the physics.

“And when we did that, it popped out that what the neural network is learning is a combination of low-pass filters, high-pass filters and Gabor filters,” he said.

“The beautiful thing about this is, the neural network is not doing any magic,” Hassanzadeh said. “It’s not doing anything crazy. It’s actually doing what a physicist or mathematician might have tried to do. Of course, without the power of neural nets, we did not know how to correctly combine these filters. But when we talk to physicists about this work, they love it. Because they are, like, ‘Oh! I know what these things are. This is what the neural network has learned. I see.’”

Subel said the findings have important implications for scientific deep learning, and even suggest that some things scientists have learned from studying machine learning in other contexts, like classification of static images, may not apply to scientific machine learning.

“We found that some of the knowledge and conclusions in the machine learning literature that were obtained from work on commercial and medical applications, for example, do not apply to many critical applications in science and engineering, such as climate change modeling,” Subel said. “This, on its own, is a major implication.”

Reference: “Explaining the physics of transfer learning in data-driven turbulence modeling” by Adam Subel, Yifei Guan, Ashesh Chattopadhyay and Pedram Hassanzadeh, 23 January 2023, PNAS Nexus.
DOI: 10.1093/pnasnexus/pgad015

Chattopadhyay received his Ph.D. in 2022 and is now a research scientist at the Palo Alto Research Center.

The research was supported by the Office of Naval Research (N00014- 20-1-2722), the National Science Foundation (2005123, 1748958) and the Schmidt Futures program. Computational resources were provided by the National Science Foundation (170020) and the National Center for Atmospheric Research (URIC0004).

READ  Los pronósticos de nubes muestran dónde el cielo será perfecto para una lluvia de meteoritos o una posible tormenta de meteoritos

Experiencia en periódicos nacionales y periódicos medianos, prensa local, periódicos estudiantiles, revistas especializadas, sitios web y blogs.

Continue Reading
Click to comment

Leave a Reply

Tu dirección de correo electrónico no será publicada.

Horoscopo

Las rocas debajo de la capa de hielo de la Antártida revelan un pasado sorprendente

Published

on

El campamento de campo en el glaciar Thwaites donde se basó el equipo para la perforación. Crédito: Greg Balco (Centro de Geocronología de Berkeley)

Investigadores internacionales[{» attribute=»»>Thwaites Glacier Collaboration found that the West Antarctic Ice Sheet had been thinner in the past and had regrown, suggesting that glacial retreat could be reversed. The study used rock samples to show that ice near Thwaites Glacier was at least 35 meters thinner in the last 5000 years and took a minimum of 3000 years to reach its current size. However, this recovery timeframe poses concerns given the expected impact of sea level rise due to imminent climate change.

The West Antarctic Ice Sheet is shrinking, with many glaciers across the region retreating and melting at an alarming rate. However, this was not always the case according to new research published last month in The Cryosphere. A team of scientists from the International Thwaites Glacier Collaboration (ITGC), including two researchers from the British Antarctic Survey (BAS), discovered that the ice sheet near Thwaites Glacier was thinner in the last few thousand years than it is today. This unexpected find shows that glaciers in the region were able to regrow following earlier shrinkage.

Sea level rise is already putting millions of people in low-lying coastal communities around the world at risk from flooding. The contribution from melting Antarctic ice is currently the greatest source of uncertainty in predictions of how much and how quickly the sea level will rise in the coming decades and centuries. Together with its immediate neighbor, Thwaites Glacier currently dominates the Antarctic contribution to sea level rise. To understand how this important glacier will respond to the climate changes expected in the coming century, scientists need to know how it behaves under a wide range of climatic conditions and over long timescales. Since satellite observations only go back a few decades in time, we need to look at the geological record to find this information.

Thwaites Rock Core

The rock cores were taken back to the lab from Thwaites for analysis. Credit: Keir Nichols (Imperial College London)

Jonathan Adams, co-author and PhD student at BAS, says:

“By studying the history of glaciers like Thwaites, we can gain valuable insight into how the Antarctic Ice Sheet may evolve in future. Records of ice sheet change from rocks that are presently exposed above the ice sheet surface end around 5000 years ago, so to find out what happened since then, we need to study rock presently buried beneath the ice sheet.”

Using drills specially designed to cut through both ice and the underlying rock, the team recovered rock samples from deep beneath the ice sheet next to Thwaites Glacier. They then measured, in those rock samples, specific atoms that are made when rocks are exposed at the surface of the Earth to radiation coming from outer space. If ice covers those rocks, these particular atoms are no longer made. Their presence can therefore reveal periods in the past when the ice sheet was smaller than the present.

Keir Nichols, a glacial geologist from Imperial College London and a lead author of the study, says:

“This was a huge team effort: several of us spent weeks away from home doing fieldwork in an extremely remote part of Antarctica, whilst others endured literally thousands of hours in the lab analyzing the rocks we collected. The atoms we measured exist only in tiny amounts in these rocks, so we were pushing right to the limit of what is currently possible and there was no guarantee it would work. We are excited that this is the first study to reveal the recent history of an ice sheet using bedrock collected from directly beneath it.”

The team discovered that the rocks they collected were not always covered by ice. Their measurements showed that, during the past 5000 years, ice near Thwaites Glacier was at least 35 meters thinner than it is now. Furthermore, their models demonstrated that its growth since then – making the ice sheet the size it is today – took at least 3000 years.
This discovery reveals that ice sheet retreat in the Thwaites Glacier region can be reversed. The challenge for scientists now is to understand the conditions required to make that possible.

Joanne Johnson, a geologist at BAS and co-author of the study, says:

“On the face of it, these results seem like good news – Thwaites Glacier was able to regrow from a smaller configuration in the geologically-recent past. However, our study shows that this recovery took more than 3000 years, in a climate that was likely not as warm as what we expect for the coming centuries. If we want to avoid the impacts of sea level rise on our world that will result from continued retreat of the West Antarctic Ice Sheet, that timescale is far longer than we can afford to wait.”

Reference: “Reversible ice sheet thinning in the Amundsen Sea Embayment during the Late Holocene” by Greg Balco, Nathan Brown, Keir Nichols, Ryan A. Venturelli, Jonathan Adams, Scott Braddock, Seth Campbell, Brent Goehring, Joanne S. Johnson, Dylan H. Rood, Klaus Wilcken, Brenda Hall and John Woodward, 28 April 2023, The Cryosphere.
DOI: 10.5194/tc-17-1787-2023

READ  El estudio ofrece evidencia nueva y más nítida de la tectónica de placas temprana y los cambios de polos geomagnéticos
Continue Reading

Horoscopo

Complejidad inesperada de estructuras misteriosas en la Vía Láctea

Published

on

Una superposición de una imagen de la Vía Láctea, tomada por el observatorio espacial Gaia de la Agencia Espacial Europea, y una visualización de simulaciones de las burbujas eRosita y Fermi. Un nuevo estudio publicado en astronomía natural proporcionó información sobre las propiedades de las burbujas de eRosita, estructuras gigantes de gas de alta energía que se extienden por encima y por debajo del centro de la galaxia de la Vía Láctea. Crédito: ESA/Gaia/DPAC

Una nueva mirada a los datos antiguos revela nuevos detalles sobre la formación galáctica.

Los astrónomos han descubierto que las burbujas eRosita, estructuras gaseosas de alta energía en el[{» attribute=»»>Milky Way, are more complex and not hotter than surrounding areas, contrary to previous assumptions. Their analysis of Suzaku satellite data suggests the bubbles originate from nuclear star-forming activity rather than a supermassive black hole.

Astronomers have revealed new evidence about the properties of the giant bubbles of high-energy gas that extend far above and below the Milky Way galaxy’s center.

In a study recently published in Nature Astronomy, a team led by scientists at The Ohio State University was able to show that the shells of these structures – dubbed “eRosita bubbles” after being found by the eRosita X-ray telescope – are more complex than previously thought.

Although they bear a striking similarity in shape to Fermi bubbles, eRosita bubbles are larger and more energetic than their counterparts. Known together as the “galactic bubbles” due to their size and location, they provide an exciting opportunity to study star formation history as well as reveal new clues about how the Milky Way came to be, said Anjali Gupta, lead author of the study and a former postdoctoral researcher at Ohio State who is now a professor of astronomy at Columbus State Community College.

These bubbles exist in the gas that surrounds galaxies, an area that is called the circumgalactic medium.

“Our goal was really to learn more about the circumgalactic medium, a place very important in understanding how our galaxy formed and evolved,” Gupta said. “A lot of the regions that we were studying happened to be in the region of the bubbles, so we wanted to see how different the bubbles are when compared to the regions which are away from the bubble.”

Previous studies had assumed that these bubbles were heated by the shock of gas as it blows outward from the galaxy, but this paper’s main findings suggest the temperature of the gas within the bubbles isn’t significantly different from the area outside of it.

“We were surprised to find that the temperature of the bubble region and out of the bubble region were the same,” said Gupta. Additionally, the study demonstrates that these bubbles are so bright because they’re filled with extremely dense gas, not because they are at hotter temperatures than the surrounding environment.

Gupta and Smita Mathur, co-author of the study and a professor of astronomy at Ohio State, did their analysis using observations made by the Suzaku satellite, a collaborative mission between NASA and the Japanese Aerospace Exploration Agency (JAXA).

By analyzing 230 archival observations made between 2005 and 2014, researchers were able to characterize the diffuse emission – the electromagnetic radiation from very low-density gas – of the galactic bubbles, as well as the other hot gases that surround them.

Although the origin of these bubbles has been debated in scientific literature, this study is the first that begins to settle it, said Mathur. As the team found an abundance of non-solar neon-oxygen and magnesium-oxygen ratios in the shells, their results strongly suggest that galactic bubbles were originally formed by nuclear star-forming activity, or the injection of energy by massive stars and other kinds of astrophysical phenomena, rather than through the activities of a supermassive black hole.

“Our data supports the theory that these bubbles are most likely formed due to intense star formation activity at the galactic center, as opposed to black hole activity occurring at the galactic center,” Mathur said. To further investigate the implications their discovery may have for other aspects of astronomy, the team hopes to use new data from other upcoming space missions to continue characterizing the properties of these bubbles, as well as work on novel ways to analyze the data they already have.

“Scientists really do need to understand the formation of the bubble structure, so by using different techniques to better our models, we’ll be able to better constrain the temperature and the emission measures that we are looking for,” said Gupta.

Reference: “Thermal and chemical properties of the eROSITA bubbles from Suzaku observations” by Anjali Gupta, Smita Mathur, Joshua Kingsbury, Sanskriti Das and Yair Krongold,1 May 2023, Nature Astronomy.
DOI: 10.1038/s41550-023-01963-5

Other co-authors were Joshua Kingsbury and Sanskriti Das of Ohio State and Yair Krongold of the National Autonomous University of Mexico. This work was supported by NASA.

READ  amoako boafo abre dot.ateliers, un espacio de arte innovador en accra diseñado por david adjaye
Continue Reading

Horoscopo

MDA se asocia con Toth Technology para la capacidad de conocimiento del dominio espacial canadiense

Published

on

Foto: BRA y Aurora (CNW Group/MDA Ltd.)

MDA trabajar con Tecnología Thoth para crear una capacidad canadiense de vigilancia por radar y conocimiento del dominio espacial (SDA) en el espacio profundo. Las compañías anunciaron el miércoles un acuerdo estratégico para combinar los servicios de datos comerciales de MDA con la tecnología de radar terrestre de Thoth para la vigilancia soberana en el espacio profundo de Canadá.

Thoth tiene una tecnología de radar terrestre llamada Earthfence que puede caracterizar objetos en órbita geosincrónica (GEO), incluida una instalación de radar en el norte de Ontario. La MDA proporcionará una herramienta de plataforma basada en la web para evaluar y almacenar datos de Earthfence, y brindará una interfaz de cliente para todas las solicitudes de datos.

Las empresas dijeron que Earthfence proporciona información más precisa que los sistemas ópticos actuales y, al trabajar juntas, las empresas desarrollarán capacidades «transformadoras» en la vigilancia del espacio profundo y SDA.

“MDA actualmente opera [Canada’s] La nave espacial Sapphire del Departamento de Defensa Nacional, el único contribuyente espacial no estadounidense a la red de vigilancia espacial de EE. UU., y con este acuerdo con Thoth, estamos bien posicionados para continuar brindando capacidades críticas de conocimiento del dominio espacial que son una parte esencial de la vigilancia espacial. y seguridad espacial”, comentó el director ejecutivo de la MDA, Mike Greenley.

READ  China utiliza Drag Sail para limpiar con éxito la basura espacial
Continue Reading

Trending