2020 Fiscal Year Research-status Report
Combining Collective and Artificial Intelligence to Understand the Early Universe
| Project/Area Number |
20K14464
|
|---|
| Research Institution | The University of Tokyo |
|---|
Principal Investigator |
HARTWIG Tilman 東京大学, 大学院理学系研究科(理学部), 助教 (00843434)
|
|---|
| Project Period (FY) |
2020-04-01 – 2023-03-31
|
| Keywords | Artificial Intelligence / Machine Learning / Milky Way / Stellar Parameters |
|---|
| Outline of Annual Research Achievements |
1. I have joined the Subaru PFS Galactic Archaeology Working group. We have weekly online meetings and I contribute to the development of the stellar parameter analysis pipeline.
2. A collaborator of mine, Dr Mohammad Mardini, has generated synthetic spectra of extremely metal-poor stars for this project. The stars cover the relevant range in temperature, surface gravity and metallicity in which we expect our future targets.
3. We have generated mock observations of metal-poor stars with the PFS Exposure Time Calculator. This project is carried out with Mohammad Mardini and Miho Ishigaki. These realistic spectra will be the basis to train a supervised machine learning model.
4. We wrote an Ancillary Proposal for the PFS cosmology survey to observe metal-poor halo stars. If the proposal is accepted, we can expect to increase the number of known extremely metal-poor stars by one order of magnitude.
5. I gave an online lecture on “Generative Deep Learning” to an interdisciplinary group of students. We have covered neural networks, variational auto encoders, recurrent neural networks, reinforcement learning, and transformers. With two of these students, I am now working on AI-related projects.
|
| Current Status of Research Progress |
Current Status of Research Progress
3: Progress in research has been slightly delayed.
Reason
The global Covid-19 pandemic impacted our project. International travel is restricted, and collaborators could not come to Tokyo as planned. We are trying to continue the discussions online as much as possible, but this way of communication is not ideal, and the progress of the project is therefore unfortunately slower than expected.
One pillar of this project should be “Collective Intelligence” in the form of public lectures and citizen science. Due to the pandemic, this was not possible until now (see future plans).
|
| Strategy for Future Research Activity |
We will finalise the grid of synthetic spectra and confirm the validity together with Evan Kirby (CalTech) and Laszlo Dobos (JHU).
The Covid-restrictions do not allow us to organise public lectures or other forms of citizen science events, as originally planned. If the Covid-situation improves this year, we plan to follow our original plan. Otherwise, we plan to host at least one data science challenge on Kaggle.com, a popular website for people who want to learn and practice AI. Formulating our machine learning problem as Kaggle challenge will have a similar impact as our original plan: 1) a diverse group of users can explore our astronomical data and learn machine learning. 2) we will have new scientific inspirations of how to optimally analyse the PFS data from citizen scientists.
|
| Causes of Carryover |
Due to the global Covid-19 pandemic, several items could not be executed as originally planned. Especially international travel was restricted and the money from this grant could not be spent on travel, international collaborations, and conferences. Moreover, we were not able to organise public lectures on AI, as originally planned.
The remaining money will be spent on: 1) Paying a freelance researcher to generate further synthetic spectra: 250,000JPY 2) Kaggle competition: 150,000JPY 3) Books hardware & Miscellaneous: 35,000JPY
|
