Oral Presentation
Exploring Galaxy Spectra by Deep Learning
Presenter: Daiki Iwasaki (Nagoya University)
The traditional spectral energy distribution (SED) fitting requires making many assumptions about observed SEDs of galaxies to estimate physical properties (e.g., dust model, stellar spectral model). SED fitting also requires sampling from a high dimensional space of model parameters that require high computational costs. Applying it to the big data observed by upcoming galaxy surveys such as JWST is unrealistic. We propose a generative model using unsupervised neural networks or variational autoencoder (VAE). There are four benefits of VAE. Firstly, VAE compresses high dimensional data to extract only the essential spectral features without assumptions. The second is outlier detection. It also helps us find unusual data artifacts, stars, or interesting galaxies for further study. The third is missing value completion and denoising. If we input data that has missing values and noise, the generative model can inpaint the missing data. We report the performance of our VAE model. Since VAE captures the nonlinear relationships, unlike the principal component analysis (PCA), VAE requires only 10 components to express the PCA reconstructed spectra with 650 components. We also can estimate the physical value of galaxies predicted by the SED fitting from a small number of latent variables. The estimation can be done in much less computation time than SED fitting and improve prediction accuracy results by not including extra information such as noise.
asiaa.sinica.edu.tw