Explaining (with) AI:

the Universe and other stuff

2023/12/06 Wed
by

Philipp Denzel contact_qr.png

Slides on my website

https://phdenzel.github.io/

talk_qr.svg

Link/QR code to the slides for later or to follow along

Projects at CAI

  • SKA project (funding: SERI/SKACH, collab: IVS & IWI):
    • Square Kilometer Array observatory & SKACH
    • generative modelling of sky simulations ("mocks")
    • my interests:
      • generative deep learning,
      • galactic evolution, dark matter…

zhaw_ska_team.jpeg

Figure 1: ZHAW's SKACH team at CSCS in Lugano

Projects at CAI

  • certAInty project (funding: Innosuisse, collab: RAI/IVS, IAMP & certX):
    • Certification scheme for AI systems
    • transparency of AI systems, regulations for AI systems
    • my interests:
      • XAI (in medicine), expanding my toolbox,
      • the "politics" of AI…

certAInty_collab.png

Outlook

How to build a universe

  • Cosmology: study of the observable Universe's
    • origins and evolution
    • large-scale structure
    • physical laws

How to build a universe

  • Astrophysics: ascertain the nature and dynamics of celestial bodies
    • galactic dynamics (most common light sources in the sky)
    • baryonic matter (stuff that interacts with light)
    • radiation (aka light)
    • dark matter (stuff that doesn't interact with light)

Unique scientific disciplines

  • problem with the scientific method
    1. phenomenon ⟶ question
    2. theory/hypothesis ⟶ predictions
    3. test in experiment?
    4. analysis ⟶ conclusion
      • publish & retest
  • computational simulations replace experiments
    • simulate the Universe…
    • what are the initial conditions?

Astronomical scales

Sun light

Alpha Centauri

Young galaxies

Cosmic epochs

The Cosmic Microwave Background

cmb3D_1e0.png

Figure 2: 2006, Credit: ESA/Planck

The Cosmic Microwave Background

The Big Bang

  • expansion of the Universe from an initial state (not from a singularity!)
    • at 10-43 sec: Planck epoch (high density/energy state), size ~ 10-35 m
    • at 10-37 sec: the gravitational force decouples, the Universe expands
      • Inflation: exponential expansion and cooling
    • at 10-32 sec: quark-gluon plasma, size ~ 1043 m
      • symmetry-breaking phase transitions cause other forces to separate
    • at 10-6 sec: baryons form, expansion and cooling continues
    • at 379'000 years: Universe becomes opaque ⟶ CMB

Mollweide projection

mollweide.jpg

Figure 3: 2006, Credit: NASA

CMB anisotropies

cmb2D_5e-4.png

Figure 4: 2006, Credit: ESA/Planck

Flagship cosmological particle simulations

pkdgrav3_full_sky.png

Figure 5: 2016, Credit: D. Potter, J. Stadel, R. Teyssier

Cosmological simulations using hydrodynamics

Fluids

\begin{align} \frac{\partial \rho}{\partial t} &+ \nabla\cdot (\rho\textbf{v})= 0 \label{eq:EulerMass} \\ \frac{\partial (\rho\textbf{v})}{\partial t} &+ \nabla\cdot (\rho(\textbf{v} \otimes \textbf{v}) + \mathbb{P}) = \rho \textbf{a} \label{eq:EulerMomentum}\\ \frac{\partial E}{\partial t} &+ \nabla \cdot (E + \mathbb{P}) \textbf{v} = \rho \textbf{a} \textbf{v} \label{eq:EulerEnergy} \end{align}

Radiation

\begin{align} \frac{1}{c}\frac{\partial I_{\nu}}{\partial t} + \hat{\textbf{n}}\cdot\nabla I_{\nu} &= j_{\nu} - \alpha_{\nu}I_{\nu} \label{eq:Radiative_transfer} \\ \frac{1}{c^{2}}\frac{\partial\textbf{F}_{\nu}}{\partial t} \,+\, \nabla\cdot\mathbb{P}_{\nu} &= - \frac{\alpha_{\nu}\textbf{F}_{\nu}}{c} \label{eq:Radiative_flux_moment} \\ \frac{\partial E_{\nu}}{\partial t} \,+\, \nabla\cdot\textbf{F}_{\nu} &= 4\pi j_{\nu}\,-\, \alpha_{\nu}cE_{\nu} \label{eq:Radiative_energy_moment} \end{align}

SPH simulations: "zoom-ins"

B-field (TNG100), Credit: IllustrisTNG

SPH simulations: isolated galaxies

pkdgrav3_full_sky.png

g1 (TNG50), Credit: IllustrisTNGs

Theory ↔ Simulations ↔ Observations

21st_century_instruments.png

Figure 6: 2023, Credit: SKAO

Radio telescopes





\begin{equation} V_{pq} = \int_{4\pi} g_{p}(r)\ B(r)\ g^{\ast}_{q}(r) e^{-\frac{2\pi}{\lambda}\langle\vec{p}-\vec{q}, \vec{r}\rangle} \text{d}\Omega \end{equation}








The Square Kilometer Array

Under construction

SKA-MPI-5.jpg

Figure 7: 2023, Credit: SKAO

Some numbers

on Proxima Centauri b
amazing_fact_1.png

exascale supercomputers
amazing_fact_2.png

amazing_fact_3.png
over 7 Pbps

amazing_fact_4.png
storing 750 PB/yr

Some facts

SKAO_fact_graphics.png

Figure 8: 2023, Credit: SKAO

Plans

🇿🇦 Meerkat National Park (150km extent)
ska_mid_map.png

🇦🇺 Murchison Observatory (75km extent)
ska_low_map.png

Indigenous communities

indigenous_people.png

indigenous_land_use_agreement.png

Member nations

SKAO_Member_Map_2023.png

Figure 9: 2023, Credit: SKAO

Switzerland joined in January 19 2022

wef_2022.jpg

Figure 10: celebrating at the WEF 2022, Credit: SKACH

SKA in Switzerland

  • leverage industry and technical partners
  • providing expertise in
    • the development of advanced receivers for dish antennas
    • precision timing and automation
    • signal processing
    • Big Data
  • contribute to the development of European SKA Regional Centre (SRC)

SKACH

group_photo_basel_23.jpg

Figure 11: Credit: SKACH

SKACH

audience_basel_23.jpg

Figure 12: Credit: SKACH

SKACH

skach_school_logos.png

Figure 13: Credit: SKACH

SKACH organization

work_packages.png

Figure 14: Credit: SKACH

Deep learning sky simulations

domains.png

Figure 15: Dataset of over 30'000 x 6 galaxy maps

Deep learning sky simulations

domains_directions.png

Figure 16: Use image domain translation models: observations (21cm) ↔ physical properties

cGANs (pix2pix or cycleGAN)

  • generator - discriminator pairs
  • learn the mapping from domain A   ⇿   B and vice versa

doge_starrynight.jpg

pix2pix schema

pix2pix_schema.png

Figure 17: Use pix2pix to generate dark matter maps from mock observations

Preliminary results

Ground truth

dm_predictions.png
Predictions from pix2pix

Future plans

Score-based generative modeling

diffusion.png

Figure 18: Score-based diffusion Song et al. (2021)

Uncertainty quantification by sampling from posterior

Credit: Ramzi et al. (2020)

Gravitational lensing

RAS PR 17/4, Credit: ESA/Hubble, NASA

  • Zurich lensed

Explanations are important

explanations_joke.png

There's an app for that

handwriting_app.png

Machine bias

machine_bias.png

People tend to anthropomorphize

algorithm_vs_humans.png

Regulations are coming…

Certification of AI systems

  • Objectives:
    • Development of a certification scheme for AI systems with specific objectives and means of compliance
    • Suite of technical and scientific methods to verify relevant properties of the AI-based system as basis for the certification scheme
    • Establish an explicit link between objectives from regulations and technical methods
    • Combination of processes and algorithmic methods

Principle-based approach to Trustworthy AI

principles_trustworthy_ai.png

EU Artificial Intelligence Act

principles_trustworthy_ai.png

EU_AI_Act.png

Risk-based approach

principles_trustworthy_ai.png

risk_levels.png

certAInty: a certification scheme for AI systems

certification_scheme.png

certAInty: a certification scheme for AI systems

certification_scheme_2.png

certAInty: a certification scheme for AI systems

certification_scheme_3.png

certAInty: a certification scheme for AI systems

certification_scheme_4.png

Use case: skin lesion classification

skin_lesions_benign_malignant.png

skin_lesion_classifier_scheme.png

Use case: skin lesion classification

  • ISO 23894 Artificial intelligence - Guidance on Risk management (4) + EASA Concept paper (2.2.1):
    • Identify stakeholders which in turn identify goals and means of increasing Transparency & Explainability
    • Doctor, patient, assessor, developer
  • ISO 24028:2020 WD Overview of trustworthiness in artificial intelligence (10.3.3):
    • The AI system should provide ex-ante and ex-post explanations, both means of explanations should be considered.
    • Local explanations of the AI system’s decision for doctor → patient, through communication of relevant image features in images
    • Global explanations for developer and assessor

Scenario: Physician

taxonomy_transparency_1.png

Scenario: Physician

taxonomy_transparency_2.png

Scenario: Physician

taxonomy_transparency_3.png

Scenario: Physician

taxonomy_transparency_4.png

Scenario: Physician

taxonomy_transparency_5.png

SHAP/Gradient-based methods

shap_saliency.png

Scenario: Assessor

taxonomy_transparency_6.png

Scenario: Assessor

taxonomy_transparency_7.png

Class maximization

class_maximization.png

Take-home message

  • Assessment and certification of AI systems:
    • There is a gap between requirements and technical methods
    • Need for innovation in linking means of compliance to processes and algorithmic methods
    • Guidelines for developers and users
    • Benchmarking of technical methods on real-world data

Contact

https://phdenzel.github.io/

talk_qr.svg

Created by phdenzel.