CPLUOS Astrophysics team meeting

Friday 1 Jul 2022, 14:00 → 15:10 Asia/Seoul

물리학과

https://uos-ac-kr.zoom.us/j/8264902652

14:00 → 14:50 COSMOLOGY

14:00 Hyeonmo's Report

Speaker: Hyeonmo Koo (University of Seoul)

Head-on Collsion of FDM/CDM Halos

• Draw snapshots: FDM vs CDM
  • Snapshots of density map of FDM/CDM halo collision for v_i = 112.776 km/s. For the last two snapshots of each v_i, the COM of each halos should be located on the dotted line if any other interactions are excluded. However, their velocities would decrease more than expected. For FDM Halos, both gravitational cooling and dynamical friction would lower their velocities after collision. For CDM Halos, only dynamical friction contributes. Thus, Δv of FDM would be larger than CDM.
  • The snapshots used 20,000 particles.

14:10 Young's report. FoF and MST algorithm

Speaker: Young Ju

FoF and MST algorithm

• Controlled data 1
  • Number of cluster: 50
  • sig = 1: separated data
  • sig = 10: diffused data
  • Aim: MGS algorithm can find clusters as other algorithms
• Density: 10 Mpc/h
• Density: 5 Mpc/h
• Dr. Sabiu and Young made MGS's Python version.

14:20 Hannah's Report

See the slides from 8p

Speaker: Hannah Jhee (University of Seoul)

Filament

• No progress

14:30 Sumi's report

Speaker: Sumi Kim (University of Seoul)

Plan for IAU Poster

• Halo catalogue + SDSS-Ⅲ DR12 CMASS + New Gramsci → 4 point correlation function + Machine learning → Cosmology parameter estimation → Compare with previous results
• Study more about 4pcf paper
  • Probing Parity-Violation with the Four-Point Correlation Function of BOSS Galaxies
  • Parity-violation has scientific data.

14:40 Dr.Sabiu's Report

Speaker: Dr Cristiano Sabiu (University of Seoul)

Deep Learning the Deep Sky: Recovering low surface brightness objects with machine learning

• New Candidates
  • We are planning to update the catalogue of low mass dwarf galaxies to create the most complete and comprehensive survey of the local universe.
  • A dataset that will be useful not only for galaxy physics but maybe also cosmology.
• Transfer Learning
  • The ImageNet dataset is a very large collection of human annotated photographs designed by academics for developing computer vision algorithms.
  • It has approximately 1 million images and 1,000 object classes.
  • The most popular CNN architectures have been trained on this dataset and their associated weights are available for download.
  • Can a machine that has been pre-trained to categorize everyday object also catagorize astronomical images?
• Transfer Learning
  • Let's take a popular model "VGG16".
  • It has 16 CNN layers.
  • Outputs 4X4X512 neurons.
  • Millions of weights.
  • We take all of the trained weights and use the output prediction as an input into a fully connected DNN (512) and finally output a Single neuron activation for 'dE'.
  • The pretrained network its not doing better than just doing our own training!
  • Let's fine-tune! Unfreeze the VGG weights and train for an additional 10 epochs.
  • We gain more than five% in accuracy in detecting dE with a network that has seen everyday objects - why?
  • Answer: Non-astronomers can easily learn to categorize galaxy images.
  • Fine-Tuning

14:50 → 15:00 Discussion