GraViT: Transfer Learning with Vision Transformers and MLP-Mixer for Strong Gravitational Lens Discovery

René Parlange, Juan C. Cuevas-Tello, Octavio Valenzuela, Omar de J. Cabrera-Rosas, Tomás Verdugo, Anupreeta More, Anton T. Jaelani

📘 arXiv preprint

Systematic comparison of neural networks used in discovering strong gravitational lenses

Anupreeta More, Raoul Canameras, Anton T. Jaelani, Yiping Shu, Yuichiro Ishida, Kenneth C. Wong, Kaiki Taro Inoue, Stefan Schuldt, Alessandro Sonnenfeld

📓 Monthly Notices of the Royal Astronomical Society (MNRAS)

---

🌌 C21 dataset: HOLISMOKES VI (Cañameras et al., 2021)

🔗 Astronomy & Astrophysics (A&A)

Training: 40,000 mocklenses and 40,000 non-lenses. Validation: 500 of each class.

Hyper Suprime-Cam (HSC) gri bands and Hubble Ultra Deep Field (HUDF) ACS and WFC3 bands

🌌 J24 dataset: SuGOHI X (Jaelani et al., 2024)

🔗 Monthly Notices of the Royal Astronomical Society (MNRAS)

18,660 lenses and 18,660 non-lens objects

Hyper Suprime-Cam (HSC) gri bands

No validation dataset (partitioned 95% for training and 5% for validation)

---

🧪 Experiments

HOLISMOKES VI (C21)

Notebook	Description
a1-C21-classification-head.ipynb	Fine-tune only the classification head
a2-C21-half.ipynb	Fine-tune half of the layers
a3-C21-all-blocks+ResNet18.ipynb	Train all layers + ResNet18 baseline

SuGOHI X (J24)

Notebook	Description
b1-J24-classification-head.ipynb	Fine-tune only the classification head
b2-J24-half.ipynb	Fine-tune half of the layers
b3-J24-all-blocks.ipynb	Train all layers

Combined (C21+J24)

Notebook	Description
c1-C21+J24-classification-head.ipynb	Fine-tune only the classification head
c2-C21+J24-half.ipynb	Fine-tune half of the layers
c3-C21+J24-all-blocks+ResNet18.ipynb	Train all transformer blocks + ResNet18 baseline

Subsampled C21 (18,660 samples)

Notebook	Description
s1-C21-18660-classification-head.ipynb	Fine-tune only the classification head
s2-C21-18660-half.ipynb	Fine-tune half of the layers
s3-C21-18660-all-blocks+ResNet18.ipynb	Train all layers + ResNet18 baseline

Inference

Notebook	Description
inference-L2.ipynb	Recall for search in L2 subset (138 lenses)