Yuan-Sen Ting (丁源森）
The Ohio State University
Can AI Dream of Unseen Galaxies?
Is Artificial Intelligence intelligent yet?
Submitting today (?)
The Current Landscape of AI x Astronomy
Most AI in Astronomy focuses on extending statistical methods
0.9
0.8
0.7
0.25
0.30
0.35
0.40
\Omega_M
\sigma_8
Power spectrum
Dark Matter Density
Growth Amplitude
Cheng, YST, Menard & Bruna + 20
Simulation
Based Inference
International Astrostatistics Association Award
or building effective brokers and optimzing workflows
Rubin Observatory
Can AI surprise us?
A hallmark of human intelligence is
out-of-domain generalizability
Now think of a mermaid with an elephant body
GPT-4o
Google's Veo 3
Good science comes from
counterfactual examples
Dustlane Early Type Galaxies
900,000 galaxies in Galaxy Zoo (from SDSS)
Citizen scientists
19,000 candidate galaxies (elliptical x dust features)
I invented Tinder
I invented Tinder
I invented Tinder
I invented Tinder
900,000 galaxies in Galaxy Zoo (from SDSS)
Citizen scientists
Baby YST
352 dustlane early-type galaxies
19,000 candidate galaxies (elliptical x dust features)
I made this plot with IDL
Shabala, YST+ 2012
Kaviraj, YST+ 2012
May I help?
Training a CNN classifier is hard with a small labeled dataset
Purity : 43%
Completeness : 43%
Ma+YST+ 2025
Generating more synthetic dustlane training data?
Galaxy type
Synthetic shift
Fine-tuning our own visual large language models
Ma+YST+ 2025
Counterfactual generation
Not an important plot
The CNN classifier improves with augmented training data
Purity : 43%
Completeness : 43%
Ma+YST+ 2025
72%
63%
444 newly discovered dustlane early-type galaxies
We made this plot with Python!
Ma+YST+ 2025
Kaviraj, YST+ 2012
Control Early-Type
This study
Ma+YST+ 2025
Kaviraj, YST+ 2012
Control
This study
We also made this plot with Python!
We found more (obscured) higher-redshift dustlane galaxies
Ma+YST+ 2025
This study
Kaviraj, YST+ 2012
Beyond dustlane early-type galaxies
A hallmark of human intelligence is
out-of-domain generalizability
Good science comes from
counterfactual examples
Now think of a mermaid with an elephant body
Can AI surprise us?
Maybe.
Willam: "Are you real?"
Angela: "If you can't tell, does it matter?"