Filippo Vicentini

Professor in Quantum Physics and AI at @Polytechnique . Developing ML-based simulation tools to tackle Quantum Physics.

[Filippo Vicentini @philipVinc]

Credit goes to Antoine who pushed this through.

This was also inspired by years-old discussions at w Riccardo @cqs_lab on sampling gradients (which we believe we can match with this at a lower cost) and Ao Chen who mysteriously mentioned it in some paper before removing it.

[Filippo Vicentini @philipVinc]

VMC Alert:We wondered "Are we sampling right?" And the answer seems "probably no". When doing #VMC on hard #quantum problems, |ψ(x)|² is optimal to get good "Energy estimates" but you need precise gradients to converge!

Turns out, you converge better by looking at |ψ(x)|ᵅ .

[Filippo Vicentini @philipVinc]

You can find a @NetKetOrg based repository, fully reproducible https://t.co/jcMVLdBBdG
and the algorithm is easy to implement anywhere.

[Filippo Vicentini @philipVinc]

Let me mention that I think this will be a great starting point of a wider effort to package and cleanup NQS dynamics simulation methods.

I'd like to collect interested parties that would like to work towards a wide package implementing all recent developments!

[Filippo Vicentini @philipVinc]

Do you want to simulate large spin systems with NQS? Yourself? Automatically?

@GravinaLuca97 has just released the accompanying code for the Systematic study of projected neural dynamics, done in collaboration with @vinniesavon .

Check it out! https://t.co/J3EbEnhY9i

[Filippo Vicentini @philipVinc]

So if you know how to use #nqs's @NetKetOrg , you can quickly launch simulations yourself for spin systems.

I'd like to congratulate Luca, that 'leads by example' by being open, transparent and ensuring total reproducibility of his results!

[Filippo Vicentini @philipVinc]

FYI: Ecole Polytechnique offers several benefits compared to other French academic positions (CNRS/Maitre de conference), including the possibility to negotiate startup packages, higher pay and reduced teaching load.

[Filippo Vicentini @philipVinc]

Full time Professorship openings: Ecole Polytechnique will recruit 2 full time assistant professors in Condensed Matter Physics and in Quantum Physics.

If you're working at the intersection of AI/ML and Quantum Physics or Material Science that's a plus!

For info, get in touch!

[Filippo Vicentini @philipVinc]

https://t.co/RCskloKCDA

https://t.co/PGw7i6Xizd

[Filippo Vicentini @philipVinc]

Full time Professorship openings: Ecole Polytechnique will recruit 2 full time assistant professors in Condensed Matter Physics and in Quantum Physics.

If you're working at the intersection of AI/ML and Quantum Physics or Material Science that's a plus!

For info, get in touch!

[Marin Bukov @MarinBukov]

hoping to see many of you at this upcoming machine learning 4 quantum matter workshop @mpi_pks in Dresden (Feb 24-28, '25):

https://t.co/J1MvHT9HTL

application *deadline* is _Nov 30_ !

[Giuseppe Carleo @gppcarleo]

Our endeavor on variational benchmarks for quantum many-body problems is now published in @ScienceMagazine!

In this large collaborative work (~30 institutions) led by @cqs_lab @EPFL_en, we established a practical, agreed-upon metric to quantify the hardness of quantum problems involving many particles (e.g. in materials).

The resulting metric, called v-score, is handy to compute and intrinsically suited for variational methods on both classical and quantum devices (e.g. Matrix Product States, Neural Quantum States, Variational Auxiliary Field Monte Carlo, Variational Quantum Eigensolver, DMFT solvers, and many others).

One surprising outcome is that this metric almost universally correlates with the error on ground state energies, for problems seemingly unrelated and for the tens of different techniques we have benchmarked.

This allows us to identify those problems and regions of interactions and parameters that are factually hard for existing many-body methods (e.g. spin liquids in some 3D geometries, Hubbard models for specific values of the coupling U etc.).

The v-score can assess progress of novel computational methods, both classical and quantum-based.

It will help shape criteria for measuring quantum computing performance, transitioning from generic, average-hardness qualifiers of computational complexity theory to physics-chemistry based hardness quantifiers for problems central to computational quantum science. In this domain, well-established heuristics exist, making any quantum-driven improvement both significant and highly valued.

Read the article here: https://t.co/uKfaK4WpZP

[Filippo Vicentini @philipVinc]

About code: it's spaghetti right now, and relies on an hacked version of @NetKetOrg extensively. We had to get the paper out (grants..) but we'll be cleaning it up and share it soon!

This project also triggered us to start thinking about NetKet 4. Stay tuned!

[Filippo Vicentini @philipVinc]

Neural Quantum States + Dynamics: Together with @GravinaLuca97 and @vinniesavon from @EPFL_en, we have studied all aspects of NQS projective methods to integrate Schroedinger eq.

Long paper but easy read. Lots of recap tables!
https://t.co/PLi7yuztJI

[Filippo Vicentini @philipVinc]

Overall, the adaptive hyperparameter tuning is THE thing that has made this possible. You know all those optimisations that explode, those that are suboptimal? Well, this never happens. The simulations are very stable.

I'm optimistic this will be THE way forward.

[Filippo Vicentini @philipVinc]

@PatrickKidger are do similar things but follow different APIs.

Speaking as high-level lib author taking as input a NN, I can if/else on the module type to eqx.partition/nnx.split but I would like for an agreed standard to signal 'frozen' parameters not to be `jax.grad` through.

[Filippo Vicentini @philipVinc]

Wishlist: As a 'higher-level' sciML library author (@NetKetOrg) , with end-users that want to use both equinox and linen/nnx I would love if @PatrickKidger and @cgarciae88 agreed on a common underlying API to filter and manipulate both their frameworks.

[Filippo Vicentini @philipVinc]

@PatrickKidger I do agree with your point. The right abstraction level could be a PyTree.

But my🍇 is about working with those pyTrees: Equinox and nnx expose similar concepts for leaf manipulation (eqx.partition/combine and nnx.split/merge) and 'lifted transforms' nnx/eqx.jit/grad which

[Filippo Vicentini @philipVinc]

@shoyer Deep in the library, I use AD to get gradients of trainable parameters (see https://t.co/zt71NROHP5 ).
It's not feasible to expose this to the user.
Right now we only support flax.

How can I support both flax, equinox and nnx to express nondiff pars, so deep in the library?

[Filippo Vicentini @philipVinc]

@shoyer It's unfortunately more complex than that.

I have a (non-pytree) class with complex logic holding the `model` and `parameters` as well as some domain-specific extra objects.

Today I store in this class the `flax model` and `parameters` and `model_state`.