Machine learning is one of the fastest growing areas of computer science used to build artificial intelligence system models. At the Heidelberg Institute of Theory (HITS), we are in a fortunate position to witness the development and use of machine learning tools in nearby areas, in various fields, and sometimes even in very different fields.
So, in my previous blog post on “Digital Baby Twins,” I zoomed in at the Cell level to simulate the first important month in newborn life with the help of AI. This time, we mean zooming out on machine learning-based tools. Or, in other words, when you get answers to a question, you never thought they would exist.
Opening new ground: Using machine learning for simulation
Meaningful questions can change the way you think about concepts and challenge the use of existing models. But if asking the right questions is the essence of good science, how do you find them?
Part of the inspiration often arises from realizing that older tools are inadequate, that they live longer than their usefulness, or that their maximum potential is not being used. A great starting point for hits from Astrophysticist and Postdoc Sebastian Trujillo Gomez and colleagues from Astroinformatics Group: hit IT specialist Bernd Doser along with team leader Kai Polsterer.
“When a large portion of AI applications in astronomy reviews literature aimed at automating tasks that humans specialize in, it allows for the classification and detection of anomalies in very large empty surveys.
But let's take a step back before we dive deeper. What does it mean when you talk about simulations of astronomy? Well, one of the most popular examples of a series of still-in-the-go astrophysics simulations was also developed by the hits by “theoretical astrophysics” group Volker Springel and his team. The calculations track the expansion of the universe, gravity to the gravitational quotation of matter itself, the movement or “fluidic mechanics” of cosmic gases, and the formation of stars and black holes.
So how can machine learning move the field forward?
Destruction of bias: Find “Unknown Unknown”
“In general, machines are great at learning to perform boring tasks very quickly. For this, many examples are needed. This makes machine learning ideal for automating many tasks that know the question that we are aiming to answer the data. However, this approach is not useful for finding new, unexpected patterns of data, as it is limited to our own Intu and as it is limited to our own Intu and as it is limited to our own Intu. The US Secretary of Defense was known as the famous “unknown unknown.”
To tackle this issue, Sebastian and his colleagues have developed new software tools to allow exploratory access to the cosmological simulations of the biggest Exuscars, such as Illustris. These tools learn compressed representations of large samples of simulated (or real) galaxies without human bias from data alone. It uses an interactive graphical interface to provide exploratory access to compressed representations, allowing users to explore simulated and real galaxies in the same intuitive similarity space.
This helps astronomers maximize their scientific breakthroughs by learning machines to unbiased, interpretable representations of complex data, from observational research to simulations. This tool automatically learns low-dimensional representations of complex objects such as galaxies in multimodal data (images, spectra, databases, simulated point clouds, etc.) and provides interactive exploratory access to any large data set using a simple graphical interface. This framework is interpretable and is designed to provide a pathway to work seamlessly across datasets, regardless of their origin, and discover “unknown unknowns.”
Mold Broken: Maximize your dataset
“Spherinator and its other related tool, a colleague hipster, work seamlessly to photograph catalogs containing millions of simulated galaxies, grouping together galaxies with similar structural features such as spiral arms, bulges, and bars, automatically placed on a spherical “map” where very different things stand out much more. Simulation and model drawbacks. ”
By addressing these technical challenges, researchers aim to enable scientists to more effectively extract valuable insights from simulation data, without being hampered by bias or computational limitations.
So, what are the next important questions to ask in this field? And where does inspiration come from? “We are inevitably entering the big data age of astrophysics in terms of both observational and simulation. We hope that our tools will help scientists squeeze out the most information from these new data sets, and will help to promote groundbreaking discoveries about the origins of galaxies and the properties of mystical dark matter and dark energy.”
More information about the Astroinformatics Group research at https://www.h-its.org/research/ain/, as well as papers by Kai Polsterer, Bernd Doser, Andreas Fehlner, and Sebastian Trujillo Gomez can be found here: https://www.aimodels.fyi/papers/arxiv/spherinator-hipster-representation-learning-unbiased-knowedge-discovery.
