How can choosing the wrong host galaxy for type Ia supernovae bias cosmology? 🌌
My new paper with the Dark Energy Survey collaboration investigates the impact of host galaxy mismatch on cosmology from 5 years of DES supernova data:
, 🧵1/n
Excited to be joining
@FlatironCCA
@FlatironInst
in the fall as a Flatiron Research Fellow!
time to say goodbye to philly, my home for basically a decade (!!), and hello to nyc 🏙️
Can we infer redshift from supernova photometry alone?
Introducing Photo-zSNthesis🌱, a deep learning-based approach for type Ia supernova redshift inference.
We show >5x improvement on real data over existing SN photo-z approaches!
, 🧵 1/5
The PLAsTiCC astronomical time series dataset is now on the Hugging Face Hub! 🤗 Now it only takes one line of code to download and integrate this dataset into your machine learning project - more details in the 🧵:
1/5
What's the best way to use unlabeled target data for unsupervised domain adaptation (UDA)?
Introducing Connect Later: pretrain on unlabeled data + apply *targeted augmentations* designed for the dist shift during fine-tuning ➡️ SoTA UDA results!
🧵👇
today, gen AI performance is surprisingly robust to new data/tasks, even beating specialized models! the secret: training on large-scale unlabeled data.
what can we as scientists learn from this?
some thoughts on robustness & the power of the unlabeled data you already have:
I’ll be speaking at the
@FlatironCCA
@FlatironInst
Cosmic Connections workshop on astrophysics x ML this week! Excited to learn about new directions in ML for cosmology, time domain science, and much more 💫
Check out our new overview of transformers/attention and a review of its applications in astronomy! Comments welcome 😊Thanks to
@BhuvJain
and Dimitrios for all their hard work!
Finally, all credit goes to the original developers of PLAsTiCC incl
@reneehlozek
@lgalbany
@emilleishida
@gsnarayan
@_sublunar_
and many others! I hope this is a step towards making ML in astro more accessible and reproducible. Feel free to reach out if you run into issues! 5/5
We show a >5x improvement on mean residuals over the widely used SN photo-z predictor, LCFIT+Z, on simulated/real SDSS + simulated LSST data!
We also find that performance on real data is slightly diminished (compared to simulations) but still much better than the baseline.
4/5
Is pretraining all you need?
We found that it sometimes fails to boost performance beyond no pretraining…
We figured out why and developed Connect Later to boost accuracy and robustness with pretraining in all cases!
9/10
Key results:
- Δw = 0.0013 with pure SNIa sample (incl CMB prior)
- 0.009 < Δw < 0.0032 with contamination from photometric classification
This is <10% of expected total Δw for DES-SN5YR! However, we should keep an eye on this with shrinking stat+syst in future surveys.
4/n
Finally, so grateful to my advisor Masao Sako and amazing collaborators in the DES-SN working group for all the help I received along the way. Can’t believe >20 people read and commented on this paper 🙏
5/n
Connect Later gets state-of-the-art results on real-world astro classification, tumor detection, and wildlife identification tasks!
From the WILDS robustness benchmark leaderboard ():
10/10
Key takeaways:
- directional light radius method for host matching works
- cuts on fitted lightcurve params decrease the mismatch rate (2.5% -> 1.7%) and severity (spread in Δz 1.1 -> 0.6)
- redshift-dependent photometric classifiers are less accurate with wrong redshift
3/n
Our supernova redshifts come from their host galaxies, so incorrect host match -> wrong supernova redshift -> biased Hubble diagram.
We designed DES5YR-like simulations to quantify the prevalence of host mismatches, their severity (Δz), and cosmology impact.
2/n
@dr_guangtou
thank you! the answer to the first question is what we were hoping to understand, as the presence of a trend would be more detrimental to cosmo estimates. spec follow-up would produce a redshift estimate but these spec zs are not as precise as galaxy spec z.
In the paper, we also describe a general framework for designing targeted augmentations. This design process provides an opportunity to naturally incorporate domain knowledge about the dataset!
4/5
Traditionally, SN redshifts come from host galaxies, but host galaxy redshifts can be unavailable, or host matching can be wrong (a teaser for another upcoming paper)!
Photo-zSNthesis provides an alternative that's independent of host galaxy info.
2/5
Photo-zSNthesis uses images made from SN lightcurves, a format that makes redshift visible (see low/med/high redshift examples below)! These images are then processed by a convolutional neural network, which predicts a full redshift PDF for each SN.
3/5
Here's how to use it - only one (real) line of code! It uses the Hugging Face Datasets library (see for installation instructions) and works with both TensorFlow and PyTorch. Much easier than writing custom data loading/preprocessing scripts!
4/5
First, some background on the PLAsTiCC (Photometric LSST Astronomical Time-series Classification Challenge) dataset: this is a set of ~3.5M simulated LSST lightcurves of 18 transient/variable sources including supernovae, AGN, etc. (see )
2/5
Some use cases I can think of: developing classification/regression ML models (e.g. classification by object type or redshift prediction), projects for physics/data science courses, unsupervised learning (e.g. lightcurve modeling)... the possibilities are endless!
3/5