Carla Dominguez |
Santa Cruz de Tenerife (EFE).- For researchers at the Instituto de Astrofísica de Canarias Marc Huertas and Andrés Asensio, artificial intelligence (AI) is a “competitive but surprisingly open” field, which represents an “enormous democratization” in a sector public researcher who lacks the basics: infrastructure for calculation.
In a meeting with EFE, Huertas assures that, with knowing a little programming, a code can be made “in an afternoon” that solves a problem “in an extraordinary way”, which means working “blindly” without validity ranges. .
The conclusions are made from the GTC meeting room of the Instituto de Astrofísica de Canarias (IAC), a space where up to 15 researchers in solar and stellar physics, cosmology, galaxies and biomedicine meet to “pollinate” and “optimize” learning AI collective.
The illusion of this group of researchers and doctoral students is corrupted when they remember what they lack: computer infrastructures that allow them to progress.
“We need places to do calculations”, says Andrés Asensio when he comments that there are difficulties in the administration to acquire machines, “by older generations”, and when they finally get them, “it is already too late”.
“They can arrive two years after requesting them, when they are already obsolete and do not allow the new techniques of the sector to be developed on them,” adds Marc Huertas.
This problem does not exist for private companies, such as Open AI, the artificial intelligence research company that has created the renowned Chat GPT conversation generator.
“Language models like that are not extraordinary from a technical point of view. They have brutal computing power”, adds the researcher.
Artificial intelligence, or ‘deep learning’, as they prefer to call it, is “a data science” in which, they say, “scientists will always be above systems”.
Astrophysicist Carlos Westendorp, one of the researchers in the group, adds that “natural intelligence is an expert in data classification,” a vital process in the development of AI that assumes that “systems are guided and learned by us.”
PhD student in solar physics Andreu Vicente confirms this reality when he accelerates the calculation processes that deduce the properties of the solar atmosphere based on observations.
“The solution that automation offers me is fast and good, but it doesn’t take into account all the physical factors involved, such as conservation of mass or energy,” he explains.
Asensio and Westendorp agree on how media reality has exaggerated the value of artificial intelligence that, in its intra-story, requires “impressive human work, from the design of the experiment to the introduction and cleaning of data.”
In fact, “the whiting that bites the tail” of these scientists is, precisely, in a scientific decision: the balanced or not presence of data in AI systems.
The consequence of the question is a biased scientific result: “If you study exceptions that occur in the Universe, and you have never taught these systems an exceptional case, it is very likely that you will not see it”, exemplifies Asensio.
The person experiencing this fact is Nataliya Ramos Chernenko, a doctoral student who works with machine learning techniques to determine the distance between galaxies.
The neural networks with which he investigates are trained with limited data, which is extrapolated and produces an inevitable bias, and astrophysical research on observing distant objects, for example, has also failed to piece together big data without bias.
Andrés Asensio adds that, “although there are ways to avoid them, some are not detected until after a while”, when the research project, of between four and five years, has culminated.
Despite the difficulties witnessed, the group boasts of tranquility in the face of the irruption of image generators created with AI.
Their scientific work creates synthetic images, which start from a physical model with parameters controlled 100 percent by themselves.
A physical model based on scientific equations that seek to “observe and understand the Universe” and with the help of machine learning techniques, even temporary obstacles to more traditional research are overcome.
The cosmological study by Francesco Sinigaglia, also a doctoral candidate in this group, is an example of this.
The researcher tries to recreate the structure of the Universe with ‘deep learning’ to forget months of simulation on supercomputers and model in minutes or even seconds, with normal computers.
The scientific team does not doubt the usefulness of machine learning techniques in fields such as tourism or climate change, but there are drawbacks: there is insufficient and classified data or it is restricted.
Despite current and future collaborations, such as the one they maintain with the University Hospital of the Canary Islands (HUC) for the early detection of colon cancer, the scientific team prefers the lines of research in their own sector, which also have an impact on life daily.
Elena G. Broock, a researcher in the Solar Physics group, studies the activity on the side of the Sun that we do not see from Earth through inferred data on the side that we do see. She does it thanks to artificial neural networks.
A line of work that seeks to “avoid or predict energy events that affect our telecommunications,” he adds. EFE