The secret life of an IAIFI Fellow

florpi

Building a space for the intersection of Physics and AI

https://florpi.github.io/

IAIFI Fellow

Carol Cuesta-Lazaro

Undergrad in Physics

Madrid (Spain)

Heidelberg (Germany)

2017: first ML project as an assistant researcher

First time can combine all my interests!

PhD in Data Intensive Science

Durham (UK)

Working on cosmology (theory)

+ Industry placements

IAIFI Committees

How I thought it would look like

How it looks like

Industry Partnership

Helped organise industry lunches

Early Career and Equity

Moderated panel on Industry x Academia collaborations (IAIFI workshop 2024)

Yearly gathering feedback from ECRs

Pushed for fellows involvement in felowship hiring

(Marisa and Thomas thinking about all the work this will really take that we are totally overlooking)

IAIFI Activities

Outreach

AstroxML Hackathon

Teaching

AstroML Hackathon 2024

LatamSummerSchool 2023

HerWILL 2024

IAIFISummerSchool2024

CambridgeScienceFestival2022

Ethics&ML2023

My research interests

AI4Science

Equivariance & Symmetries

Anomaly detection

Out-of-Distribution

Interpretability

Quantifying Uncertainties

Partial Observations

PDEs

Multimodal

Simulation-based-Inference

Foundation Models

Hierarchical

IAIFI Coffee

Hanging out at the penthouse

IAIFI Seminars

Hackathons

["A point cloud approach to generative modeling for galaxy surveys at the field level" 
Cuesta-Lazaro and Mishra-Sharma 

arXiv:2311.17141]

Base Distribution

Target Distribution

  • Sample
  • Evaluate

Siddharth Mishra-Sharma

IAIFI Fellow

Modelling Galaxy Point Clouds

ICML AstroML Spotlight 2023

Fixed Initial Conditions / Varying Cosmology

p(\theta|x) = \frac{p(x|\theta)p(\theta)}{p(x)}

Diffusion model

CNN

Diffusion

Increasing Noise

p(\sigma_8|\delta_m)
p(\sigma_8|\delta_m + 0.01 \epsilon)
p(\sigma_8|\delta_m + 0.02 \epsilon)
["Diffusion-HMC: Parameter Inference with Diffusion Model driven Hamiltonian Monte Carlo" 
Mudur, Cuesta-Lazaro and Finkbeiner]

 

Nayantara Mudur

CNN

Diffusion

Douglas Finkbeiner

IAIFI JI

IAIFI SI

NeurIPs ML for the Physical Sciences 2023

All learnable functions

All learnable functions constrained by your data

All Equivariant functions

More data efficient!

Incorporating symmetries in ML architectures

Non-equivariant

Julia Balla

Siddharth Mishra-Sharma

> x2 in constraining power

From Angstroms to Gigaparsecs

Tess Smidt

IAIFI JI

IAIFI Fellow

IAIFI SI

p_\phi(\rho_\mathrm{DM}|\rho_\mathrm{Galaxies})

1 to Many:

["Debiasing with Diffusion: Probabilistic reconstruction of Dark Matter fields from galaxies" 
Ono et al arXiv:2403.10648]

 

Victoria Ono

Core Park

Truth DM

Inferred DM

Observed

Nayantara Mudur

Probabilistic dark matter mapping

NeurIPs Physical Sciences 2023

IAIFI JI

Harvard students

["3D Reconstruction of Dark Matter Fields with Diffusion Models: Towards Application to Galaxy Surveys" 
Park, Mudur, Cuesta-Lazaro et al (in-prep)]

 

Posterior Sample

Posterior Mean

Debiasing Cosmic Flows

ICML AI4Science 2024

Rewinding the Universe

Stochastic Interpolants

NF

p(\delta_\mathrm{ICs}, \theta|\delta_\mathrm{Final}) =
p(\delta_\mathrm{ICs}|\delta_\mathrm{Final})
p(\theta|\delta_\mathrm{ICs},\delta_\mathrm{Final})

Michael Albergo

IAIFI Fellow

Daniel Eisenstein

IAIFI SI

Simulating what you need

(and sometimes what you want)

Compressing cosmological simulations

Brandon Y. Feng

(in Bill Freeman's group)

IAIFI SI

Ge Yang talking about NERFs at IAIFI colloquium

~ 10 trillion particles per snapshot stored

x Discrete snapshots

Can we learn compressed continuous representations with Neural Fields?

IAIFI Coffee

Learning priors from simulations

Mikhail Ivanov

IAIFI Affiliate

Why IAIFI works

Bring postdocs excited about AI+Physics

+ get them great junior collaborators

+ gpus

=

New ideas, best practices and state of the art ML

+ resources

New community at the intersection

summer schools, hackathons, workshops

deck

By carol cuesta

deck

  • 63