Automatic physical inference with information maximising neural - - PowerPoint PPT Presentation

▶

Jun 05, 2023 15 likes •77 views

Automatic physical inference with information maximising neural networks Physical Review D 97 , 083004 arXiv:1802.03537 github:information_maximiser Tom Charnock Institut dAstrophysique de Paris charnock@iap.fr DOI 10.5281/zenodo.1175196

SLIDE 1

Automatic physical inference with information maximising neural networks

Physical Review D 97, 083004 arXiv:1802.03537

DOI DOI 10.5281/zenodo.1175196 10.5281/zenodo.1175196

github:information_maximiser

Tom Charnock Institut d’Astrophysique de Paris charnock@iap.fr

SLIDE 2

How would we like to do inference?

410 420 430 440 450 460 [nm] 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 Simulated quasar flux [Photon counts]

ℒ(d|𝜄) 𝒬 (𝜄|d) = ℒ(d|𝜄)𝓆(𝜄) 𝓆(d)

4 2 2 4 Amplitude of scalar perturbation scaling ln ( |d)

Tom Charnock charnock@iap.fr Physical Review D 97, 083004, arXiv:1802.03537,

DOI DOI 10.5281/zenodo.1175196 10.5281/zenodo.1175196

SLIDE 3

How can we do inference without a likelihood?

410 420 430 440 450 460 [nm] 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 Simulated quasar flux [Photon counts]

{𝜄|𝜄 ↶ 𝓆(𝜄)}

410 420 430 440 450 460 [nm] 5 10 15 20 Simulated quasar flux [Photon counts]

𝒼 ∶ d → 𝓎

4 2 2 4 Amplitude of scalar perturabtion scaling ln 30 20 10 10 20 30 Compressed summary 4 2 2 4 Amplitude of scalar perturbation scaling ln ( |x)

Tom Charnock charnock@iap.fr Physical Review D 97, 083004, arXiv:1802.03537,

DOI DOI 10.5281/zenodo.1175196 10.5281/zenodo.1175196

SLIDE 4

How can we fjnd the summaries?

Tom Charnock charnock@iap.fr Physical Review D 97, 083004, arXiv:1802.03537,

DOI DOI 10.5281/zenodo.1175196 10.5281/zenodo.1175196

SLIDE 5

Automatic physical inference with information maximising neural networks

▶ Simulate data at a fjducial parameter value ▶ Train the network to increase the Fisher information ▶ Compress real data using trained network ▶ Do ABC with the optimally compressed summaries

Tom Charnock charnock@iap.fr Physical Review D 97, 083004, arXiv:1802.03537,

DOI DOI 10.5281/zenodo.1175196 10.5281/zenodo.1175196

SLIDE 6