1 of 33

Licia Verde

Instituto de ciencias del cosmos, Universitat de Barcelona

Closing remarks

2 of 33

The view from the tail, no, not that tail, the other tail

Licia Verde

Closing remarks

3 of 33

Back to 2021

Closing remarks

4 of 33

5 of 33

I set up a trap: I made the audience vote (anonymously)

Blue: traditional analysis

Purple: ML

Black +: LCDM prediction

Hypothetical scenario 10 years from now, axes are two parameters (you choose)

6 of 33

Should the acceptance depend on the agreement of the finding with pre-conceptions (expectations)?

NO

The results of the poll indicated the contrary

does/will the acceptance depend on the agreement of the finding with pre-conceptions? 

��

Licia Verde

Closing remarks

7 of 33

In 2023 edition language�Beware of confirmation bias!

Licia Verde

Closing remarks

8 of 33

9 of 33

Now in 2023 do we add other entries?

10 of 33

11 of 33

4.5 h of debates

8 h review talks

12 h talks

Striking: low average age of participants

A new scientific truth does not triumph by convincing its opponents and making them see the light,

but rather because its opponents eventually die and a new generation grows up that is familiar with it ...

The Planck principle

5 days

Licia Verde

Closing remarks

12 of 33

Some considerations….

Licia Verde

Closing remarks

13 of 33

Two years later… �much much more awareness, maturity

  • While not a coherent “program” a lot of work has been done.

Dunning-Kruger effect

Licia Verde

Closing remarks

14 of 33

  • A “Snowmass”?

Licia Verde

Closing remarks

15 of 33

Some considerations….

Licia Verde

Closing remarks

16 of 33

On data…

  • Simulations are not data

​

  • There are mock data (end-to-end), there are idealized data (model-generated data), and there are data-data (“real” data).

ML thrives on big data (training, making “sense” of, sifting through,…). “complexity”

These are NOT the same thing, they are PROFOUNDLY different

Proposal: m-data, i-data, data

Licia Verde

Closing remarks

17 of 33

On priors (or biases?)

  • The traning (m-,i-)data → in-built prior

​

​

​

  • The architecture → in-built prior

This is not good or bad, but it is there!

Licia Verde

Closing remarks

18 of 33

On.. The Universe

​

  • We only have one Universe. (d’oh)
  • Complexity (see opening talk)
  • Not all astronomy contexts/applications deal with this in the same way: repeated observations, contexts when confirmation/follow up is the aim, [planets, transits,…], photo-z vs cosmological parameters

ML thrives on big data (training, making “sense” of, sifting through,…)

It all depend on the question one is trying to answer.

It is very important to specifiy extremely well what is the question and why the answer is of value.

Licia Verde

Closing remarks

19 of 33

On “truth” (see dabate #1,2)

  • we are after fundamental physics (and we look up at the sky to find it)
  • Physical law vs symmetries vs conceptual framework (GR, QM) vs theory (string theory, inflation…) vs model (LCDM, wCDM...) vs effective model (cz=H0 d) vs empirical relation (PL relation, Phillps relation)

​

(not all models are created equal)

​

Understanding is not describing

​

Fitting cosmological parameters is not understanding

Licia Verde

Closing remarks

20 of 33

On epicycles….

Geocentric model

Heliocentric model

Kelper, Newton

Fundamental principles

But this worked well enough!

Licia Verde

21 of 33

On… different contexts

  • Discover the physics
  • Know the physics
  • Have a fiducial model
  • Have no clue and no fundamental principle model but a) don’t care b) still have to deal with it
  • Know it all but want to be fast/cheap
  • Summarize/search…

“the ML cog”

Licia Verde

Closing remarks

22 of 33

On new results… (or opening talk question)

ML-enabled

ML-enhanced

Faster, cheaper

Disruptive

otherwise impossible

Think about internal combustion engine

Is it “just” a tool?

Think about the internet

Licia Verde

Closing remarks

23 of 33

ML-enhanced

ML-enabled

Clearing the path

towards the truth

Truth= finding or

  • funtamental,

physics

Discover the truth

Licia Verde

Closing remarks

24 of 33

interpretability

on

Licia Verde

Closing remarks

25 of 33

Making the black box transparent

  • But not only: reducing dimensionality, reducing complexity, connection to the Fisher information matrix.
  • Truth in latent space (summarized in debate #1)
  • Symbolic regression….
  • Contrast learning
  • Response…. Saliency maps, sensitivity maps

​

“I want to believe”

Combination of approaches…

Licia Verde

Closing remarks

26 of 33

Shaping the box and its content

  • Geometric deep learning
  • PINN (effort going on at home)
  • …..

​

​

​

Hard code in

Or

In the loss function

Combination of approaches…

Licia Verde

Closing remarks

27 of 33

..”attention is all you need”…

“Produce new measurements”

DEEPthink disappoining answer

Licia Verde

Closing remarks

28 of 33

Detecting neutrino masses

Say we detect Mv=0.095eV from ….. A galaxy survey

Say we detect Mv=0.06eV from ….. A galaxy survey

​

Licia Verde

Closing remarks

29 of 33

Walks like a duck, looks, like a duck, smells like a duck,

but I need some more quacking tests

If it looks like a duck and quacks like a duck but it needs batteries, you probably have the wrong abstraction

Liskov principle

Closing remarks

30 of 33

features

With a direct connection to physics (possibly fundamental)

Predictability of other features...

Which have a direct connection to other aspects of the physics

Consistency tests

null tests

Robust, consistent across different analyses

Generic, not only for ML, but easier to do for more traditional approaches

Licia Verde

A question of time?

Closing remarks

31 of 33

It changes the type of valuable skills

Forces a re-evaluation of the values and what is of value

Acceleration

Efficiency

Freeing up time/resources for things that ML can’t do well

Judea Pearl:” Current machine learning systems operate, almost exclusively, in a statistical,

or model-free mode , which entails severe theoretical limits on their power and performance.”..

humans can imagine alternative hypothetical environments for planning and learning.”casuality”,

“counterfactuals” , “what if”. Current algorithms lack causal reasoning.

​

​

Licia Verde

Closing remarks

32 of 33

ML is my co-pilot

But… where do you want to go?

33 of 33

Thank you

  • To all the speakers, panelists, in Paris/NY.

I learned a lot I can’t believe it’s already over.

  • To the session chairs. Impeccable and we were quite on time!
  • To all the participants in all the different timezones, for the lively questions and participation.
  • SOC and LOC and the support staff
  • and… to the organizers: only few years ago we would not even have imagined possible, but it went flawlessly, which is amazing!

​

“It can go wrong and you do not know (why)”

Licia Verde

Closing remarks