About this Research Topic
Neural Networks have created an incredible enthusiasm around the world and astrophysics has jumped into this fascinating adventure. After many developments, applications, successes and failures, Deep Learning in astrophysics seems to be mature enough to review its real power and limitations, as illustrated by several conferences held in 2021 and 2022 on this question.
Deep Learning is one of several supervised Machine Learning approaches, thus requiring a training set. These kinds of approaches have the appealing power of making predictions based on our current knowledge and seem very efficient at classifying images. The major requirement is the availability of a training set as large as possible to guarantee a reliable training of the algorithm. This is an issue in many cases, especially in astrophysics. In addition, it is difficult to avoid biases in the completeness and representativeness of the training set. Generative neural networks can increase the size of the training sets by creating artificial data. Is this scientifically ethical and acceptable? And does it really alleviate the bias problem?
The main concern about Deep Learning is the lack of mathematical understanding of how it works. It makes it difficult to interpret the results and physicists may be uneasy in trusting them. Interpretable Artificial Intelligence is an active research domain, but can it help physicists and astronomers? Validating techniques are being developed to give confidence in the analyses. What are they and how should we use them? Are they sufficient?
Finally, designing a neural network is a complicated task. Its structure depends on many choices that are most often optimized for a specific application. Using it outside its initial purpose may not be a good idea. What are the safeguards against using a black box from the shelves?
Machine Learning and Deep Learning have reached an incredible level of sophistication, for instance mixing up supervised and unsupervised techniques (semi-supervised learning), or developing self-learning or transfer learning. Deep Learning has an obvious power at predicting and can be used in numerical simulations to take into account many additional physical ingredients and equations without impacting too much the computation time and memory usage especially in multi-scale contexts. But is Deep Learning really as useful and trustworthy in inference problems?
There are many other supervised Machine Learning techniques which are both well established and mathematically clear: Support Vector Machine (SVM), Random Forests, k-Nearest Neighbors, regression analyses... Being generally safer, they can also be more efficient and reliable in certain cases. Despite being less glamorous, they should not be forgotten.
In this Research Topic, I wish to gather contributions that would help build a panorama of which techniques should be used and advised, or not, for a given purpose, with their clear limitations. This could be of considerable help to young astronomers that will unavoidably have to use Machine Learning and Artificial Intelligence at some point in their career. Contributions can be proposed in different formats: Original Research, Brief Research Report, Review and Mini Review, General Commentary, Hypothesis and Theory, Opinion and Perspective.
The Specialty Chief Editors of Frontiers in Astronomy and Space Sciences launch a new series of Research Topics to highlight current challenges across the fields of Astronomy and Space Sciences. Other titles in the series are:
Editor's Challenge in Astronomical Instrumentation: Machine Learning Advances
Editor's Challenge in Exoplanets: Next Generation of Exoplanet Research
Deep Learning is one of several supervised Machine Learning approaches, thus requiring a training set. These kinds of approaches have the appealing power of making predictions based on our current knowledge and seem very efficient at classifying images. The major requirement is the availability of a training set as large as possible to guarantee a reliable training of the algorithm. This is an issue in many cases, especially in astrophysics. In addition, it is difficult to avoid biases in the completeness and representativeness of the training set. Generative neural networks can increase the size of the training sets by creating artificial data. Is this scientifically ethical and acceptable? And does it really alleviate the bias problem?
The main concern about Deep Learning is the lack of mathematical understanding of how it works. It makes it difficult to interpret the results and physicists may be uneasy in trusting them. Interpretable Artificial Intelligence is an active research domain, but can it help physicists and astronomers? Validating techniques are being developed to give confidence in the analyses. What are they and how should we use them? Are they sufficient?
Finally, designing a neural network is a complicated task. Its structure depends on many choices that are most often optimized for a specific application. Using it outside its initial purpose may not be a good idea. What are the safeguards against using a black box from the shelves?
Machine Learning and Deep Learning have reached an incredible level of sophistication, for instance mixing up supervised and unsupervised techniques (semi-supervised learning), or developing self-learning or transfer learning. Deep Learning has an obvious power at predicting and can be used in numerical simulations to take into account many additional physical ingredients and equations without impacting too much the computation time and memory usage especially in multi-scale contexts. But is Deep Learning really as useful and trustworthy in inference problems?
There are many other supervised Machine Learning techniques which are both well established and mathematically clear: Support Vector Machine (SVM), Random Forests, k-Nearest Neighbors, regression analyses... Being generally safer, they can also be more efficient and reliable in certain cases. Despite being less glamorous, they should not be forgotten.
In this Research Topic, I wish to gather contributions that would help build a panorama of which techniques should be used and advised, or not, for a given purpose, with their clear limitations. This could be of considerable help to young astronomers that will unavoidably have to use Machine Learning and Artificial Intelligence at some point in their career. Contributions can be proposed in different formats: Original Research, Brief Research Report, Review and Mini Review, General Commentary, Hypothesis and Theory, Opinion and Perspective.
The Specialty Chief Editors of Frontiers in Astronomy and Space Sciences launch a new series of Research Topics to highlight current challenges across the fields of Astronomy and Space Sciences. Other titles in the series are:
Editor's Challenge in Astronomical Instrumentation: Machine Learning Advances
Editor's Challenge in Exoplanets: Next Generation of Exoplanet Research
Keywords: Astrophysics, AI methods power and limits, Deep Learning
Important Note: All contributions to this Research Topic must be within the scope of the section and journal to which they are submitted, as defined in their mission statements. Frontiers reserves the right to guide an out-of-scope manuscript to a more suitable section or journal at any stage of peer review.
