Next generation astrophysics in the era of digital intelligence

Astrophysics is on the verge of being transformed into a science that is truly data-driven, thanks to a new generation of telescopes that will deliver unprecedented volumes of data in the coming years. The analysis of these data is a major challenge and overcoming it will result in breakthrough discoveries in astrophysics and fundamental physics.

With Montreal’s leading expertise in the field of digital intelligence and its strong astrophysics community, we are in a unique position to become a major player in this field, and thus a rallying point for Canadian and global communities.

The strategic theme we are therefore proposing aims to produce major discoveries in astrophysics by utilizing new methods of data analysis and computation. To accomplish this, we have built a multidisciplinary team of outstanding researchers who are committed to this goal.

So why invest in the field of astrophysics?

Astrophysics has grown very rapidly in recent decades. For example, in the past 9 years, 4 Nobel Prizes have been given to discoveries in astrophysics. Out of 10 Nobel Prizes awarded in astrophysics over the past 120 years, 6 have been awarded in the past 20 years. Many universities have built new departments, centers and institutes in astrophysics. Large amounts of private and governmental funds have been invested in this area (e.g. Kavli institutes of Cambridge, Stanford, Chicago, Perimeter Institute, etc.). One of the major reasons for all the progress in this field is a broad public interest at all levels of society, from elementary students who are fascinated by images of space to international private foundations who heavily invest in this field.

Several new telescopes and observatories have or will start operating in the next few years. Most are multi-billion dollar international projects, which will yield data with unprecedented quality for major discoveries. This is causing and will continue to cause a major acceleration in the field.

Canada has already invested a significant amount of funds in the construction of these observatories, but to date there is still very little that is in place to analyze the data. As Canadians, we will have access to data products worth billions of dollars in investment in infrastructure. With a minor investment in data analysis, we will be able to get the maximum benefit from this data and reap as many rewards as possible.

Our project is a truly interdisciplinary collaboration of experts in machine learning, statistics and astrophysics. With the unique expertise in digital intelligence in Montreal, a true structure of interdisciplinary collaboration in this field will be an exceptional project that will produce spectacular results. This theme will generate several other impacts, notably on the training and education of students and HQP, a priority for provincial and national government agencies. We plan to implement a significant outreach component, and are also considering industrial partnerships, internship programs, visitor programs, etc. We believe that this will bring IVADO and our institutions to an entirely new level of visibility and global leadership.

Why should we act now?

The Canadian Astronomy Long Range Plan 2020-2030 (https://casca.ca/?page_id=75) clearly states that the unprecedented volumes of data in the coming years will be a major challenge for the field, but also science in general given the unique challenges astrophysical data bring to data-driven fields. Building digital research infrastructures, at the intersection of digital intelligence and astrophysics, has therefore become one of the top priorities for national and international institutions. In other words, if Montreal does not take the initiative to do this now, other institutions will.

Yet, Montreal has a clear advantage over other leading institutions: IVADO and affiliated institutions have proven that Montreal is a global center for digital intelligence (machine learning, deep learning, artificial intelligence, etc). This places Montreal in a uniquely strategic position to transform the field of astrophysics by developing new cutting edge digital intelligence tools that will truly transform astrophysics in an innovative way.

(ajout 22/07) What novel aspects can astrophysics – compared to other scientific fields – bring to the field of digital intelligence? In other words, how can investing in a fundamental science such as astrophysics enable breakthroughs in digital intelligence?

As argued above, the goal of this theme is to make breakthrough discoveries that will revolutionize our understanding of the Universe. It is based on a truly multidisciplinary approach.

However, a crucial point is that astrophysics – due to the immense quantity and complex nature of its data – also offers a unique playground for computer scientists.

More specifically:

1) The quantity of data available in astrophysics from many different surveys, telescopes, simulations, etc. is truly colossal.

The examples of new observatories and surveys are numerous: Euclid, LSST, WFIRST, DES, DESI, CHIME, Gaia, the Simons Observatory, CMB-S4 experiments, JWST, SKA, etc. To pick a specific example, the Vera Rubin Observatory, which will conduct the Legacy Survey of Space and Time (LSST) starting in 2023, will scan the entire Southern sky every 3 nights, collecting over 30 Tb of data every night for the 10 years of its operation. The Canadian community, including the core members of this proposed theme, will have access to the data produced by these experiments. The large volumes of data can provide many different learning opportunities but also advance machine learning models by pushing them to their limits in terms of dealing with extremely large datasets.

This makes the field of astrophysics the next clear data-driven field for computer scientists and machine learners.

2) The nature of astrophysical data is also complex, and comes with its unique set of challenges. This will demand the development of innovative methods in digital intelligence, which in turn will impact numerous other fields.

a) To name a specific example, astrophysical signals of interest are often extremely faint and buried under very high levels of noise. This requires robust statistical inference in order to extract the information of interest from data, and, most importantly, robust quantification of uncertainties. This (the quantification of the uncertainties associated with the outputs of machine learning methods) is a domain which has been explored – to an extent – within the context of other fields of application, but for which astrophysics is the perfect playground for a systematic and sound development and comparison to traditional Bayesian methods.

The ability to produce accurate uncertainty estimates for a variety of machine learning methods also has direct socio-economic benefits. For example, techniques that have been used to analyze astrophysical images have been (and will continue to be) of great interest to the medical community. This includes cancer screenings, where beyond quantifying the certainty with which a mass is a malignant tumor or not has obvious benefits, but also identifying zones where a reconstructed image of an MRI scan is less certain. Such techniques can also be directly applied to other fields such as self-driving cars (which can benefit greatly from accurate uncertainty of the image segmentation they perform).

b) The heterogeneity of astrophysical data (e.g., observation of multiple telescopes from the same regions, different forms of data, etc.) also provides unlimited opportunities for developing and studying generalisation, domain adaptation, etc., one of the key challenges of modern machine learning methods.
c) In astrophysics, it is possible to obtain labeled data quite systematically and accurately, by using ab initio simulations. In other fields (e.g., medical imaging), obtaining ground truth and labeled data could be a major impediment. But in astrophysics, detailed ab initio simulations are commonly used to study systems in depth which can often be used to train models and produce training data. These simulations also provide an opportunity for advancing generative models. In addition to discriminative data analysis models, generative models also are of high value in astrophysics and detailed simulations could be used for training them.

d) One of the key goals of astrophysics is using models to perform inference, which typically involves extremely complex optimisation tasks. This falls inline with the general umbrella of optimisation and operational research. The novel methods developed for astrophysics here will be readily applicable to other generic optimisation problems.

e) One of the most challenging and promising ways of exploring astrophysical data is unsupervised learning with the goal of discovering truly surprising and previously unknown physical phenomena. This is at the cutting-edge of machine learning and can push truly novel and intelligent machine learning methods to be developed.

These are just a few examples that illustrate why astrophysics – due to the quantity and nature of its data – will enable the development of next generation digital intelligence. We therefore anticipate that this theme, although focusing on the fundamental science that is astrophysics, will have major socio-economic repercussions across a broad range of fields.

Another direct consequence is the HQP that will have been trained through this theme. Physics is not a career oriented field, as opposed to other fields such as law or medicine in which students are trained to do a specific job (lawyer/medical doctor). Physics on the other hand trains its students such that they become scientists that can tackle a variety of complex problems. As a result, physicists are highly sought out on the job market because they have a unique set of skills that can be used in all fields (ex. banking, outreach, start-ups, video games, new generation of detectors, etc). We anticipate that this theme will produce the next generation of scientists, specialized in machine learning that will go onto fruitful careers in the industry, government labs, etc.

As an overall summary: there are currently no other initiatives within Canada that aim to bridge the gap between astrophysics and digital intelligence. This theme therefore aims to be the first of its kind in Canada. With Montreal’s unique expertise in digital intelligence and astrophysics, we have the potential to become a top player at the international level. However, we must act now: considering the acceleration of the field and the realization that such major breakthroughs are possible, it is only a matter of time before other universities and institutions build such an initiative.

Mots-clefs :

• Knowledge themes: Big data, Machine learning, Statistical analysis, Computational analysis, Fundamental research, Cosmology, Dark matter, Dark energy, Black Holes, Galaxies
• Talent themes: Multidisciplinary, Multi-level talent development, Mobilization of talent, EDI initiatives

Organisations pertinentes :

Several international partners have been identified:

Schmidt foundation, via Schmidt Futures donation (https://tsffoundation.org), Simons foundation via awarded grant from Simons Collaborations in Mathematics and the Physical Sciences program (https://www.simonsfoundation.org), Flatiron Institute in New York City (https://www.simonsfoundation.org/flatiron/).

Personnes pertinentes suggérées durant la consultation : 

Les noms suivants ont été proposés par la communauté et les personnes mentionnées ci-dessous ont accepté d’afficher publiquement leur nom. Notez cependant que tous les noms des professeur.e.s (qu’ils soient affichés publiquement ou non sur notre site web) seront transmis au comité conseil pour l’étape d’identification et de sélection des thèmes stratégiques. Notez également que les personnes identifiées durant l’étape de consultation n’ont pas la garantie de recevoir une partie du financement. Cette étape sert avant tout à présenter un panorama du domaine, incluant les personnes pertinentes et non à monter des équipes pour les programmes-cadres.

• Julie Hlavacek-Larrondo
• Yashar Hezaveh
• Laurence Perreault-Levasseur
• Siamak Ravanbakhsh
• Laurent Charlin
• Derek Nowrouzezahrai

We propose to pursue a very targeted strategy to achieve concrete results in a realistic amount of time. Our strategy is to identify a series of high impact problems in astrophysics where innovations in data analysis methods can produce major discoveries. The specific challenges astrophysics offers will also enable our team to produce the next generation of AI tools. We will focus most of our efforts on these specifically targeted problems. Examples include:

• The Nature and the distribution of dark matter in the Universe
• The cause of the accelerated expansion of the Universe
• The role of supermassive black holes in the Universe

These are important issues with clear avenues for their resolution.

We have a detailed plan for the strategy and activities related to this theme for the next 5 years:

1) In the first year, we will define these large projects in detail and invite specific key members who are committed to contributing to these issues. We will organize a series of internal events to launch these collaborations. We will also seek broader funding opportunities.

2) During the second year, new talents (post-docs / students) start their functions and we will continue to work collaboratively on these issues. Given that we will already have postdocs and students who have already (in 2020) started working on some of these issues, it is highly likely that we will already be able to produce results and publish a series of articles. We are going to launch an outreach program to publicize the results through our connections with the media (e.g. Les années lumière Radio Canada). We also aim to organize a conference to bring the best experts in the field to Montreal. It will promote IVADO and our institutions and create opportunities for new collaborations with other major poles. We will also build a visitor program and seminar series to continually attract new researchers and connect our initiative with the international community.

3) In the third year, we aim to have a series of important joint publications. We will form an internship program with some industrial partners and explore other industrial partnership opportunities. We already have links within Google, Microsoft and Ubisoft. It will help us place students in advantageous and competitive positions in the industry.

4) and 5) Our goal is to produce major scientific discoveries by the fifth year.

Although the specific results are of course not known, it is reasonable to think that over a mandate of 5 years with such a high level of interdisciplinary collaborations and connections with the scientific community concerned, we will be able to produce groundbreaking discoveries. Beyond the fundamental scientific problems, this theme will lead to other knowledge in digital intelligence.

After the first 5 years, our plan will be to expand the initiative to reach approximately 10 core members, 15 affiliate members, 4 postdoctoral fellows, several other postdoctoral fellows hired on specific grants and around 50 students.

(pas de documentation complémentaire pour le moment)