Speaker
Description
We present a foundation model for strong lensing based on diffusion transformers, which can solve a wide range of different simulation-based inference problems via representation learning. Image-like data such as observations are split into smaller patches, which together with their positions are embedded as tokens. Similarly, parametric models are also encoded via tokens. A strong lens system is therefore fully represented by a sequence of tokens. Each token in this sequence can be masked and the diffusion transformer is trained to infer masked tokens. Different types of inference problems such as source reconstruction, forward modeling or lens parameter estimation correspond to specific token masks. The representation as a sequence of tokens offers great flexibility. Our diffusion transformer can be trained on observations coming from different measurement instruments with different resolutions, pixel sizes or point spread functions. At the same time, both pixelated and parametric models can be used interchangeably within the same framework.
| Abstract title | A Transformer-based Foundation Model for Strong Lensing |
|---|