acerbiluigi - Network

Folks, Mastodon didn't reach critical mass for my personal usage, and I will probably discontinue this account at some point.I am trying...

Folks, Mastodon didn't reach critical mass for my personal usage, and I will probably discontinue this account at some point.

I am trying Bluesky out, which seems promising and growing very fast in the past couple of weeks (wonder why).

You can find me there at: https://bsky.app/profile/lacerbi.bsky.social

📯I am hiring again!Postdoc or PhD position in Sample-Efficient Probabilistic Machine Learning at the University of Helsinki with strong...

📯I am hiring again!

Postdoc or PhD position in Sample-Efficient Probabilistic Machine Learning at the University of Helsinki with strong links to @FCAI

Please see blurb below, and full ad here:
https://helsinki.fi/en/researchgroups/machine-and-human-intelligence/postdocphd-position-in-sample-efficient-probabilistic-machine-learning

Main goal: Extend our VBMC framework for efficient probabilistic inference (https://github.com/acerbilab/pyvbmc), with applications and collaborators in comp neuro, e.g. @internationalbrainlab

Applications evaluated on a rolling basis. Please RT!

Hi folks, I have multiple openings for PhD/postdoc positions in probabilistic machine learning at the University of Helsinki and @FCAI...

Hi folks, I have multiple openings for PhD/postdoc positions in probabilistic machine learning at the University of Helsinki and @FCAI
starting later in the year!

For more info see: https://www.helsinki.fi/en/researchgroups/machine-and-human-intelligence/open-positions

Please boost!

11/ PS: We have other tools coming out soon, follow me or our spaces for more info: https://github.com/acerbilab

11/ PS: We have other tools coming out soon, follow me or our spaces for more info: https://github.com/acerbilab

10/ We are very interested in which kind of problems you might apply PyVBMC to, and connecting to the broader open source and probabilistic...

10/ We are very interested in which kind of problems you might apply PyVBMC to, and connecting to the broader open source and probabilistic programming language crowd (e.g., @pymc, @mcmc_stan, @ArviZ).

Please get in touch if you have any questions or feedback for us!

9/ Work on this package is a joint effort of many fantastic machine learning developers working in my group at various times: Bobby Huggins,...

9/ Work on this package is a joint effort of many fantastic machine learning developers working in my group at various times: Bobby Huggins, Chengkun Li, Marlon Tobaben, Mikko Aarnos.

Thanks to @FCAI & UnivHelsinkiCS for funding and supporting the project!

8/ References:Our original VBMC papers, detailing the algorithm:- NeurIPS 2018: https://arxiv.org/abs/1810.05558- NeurIPS 2020:...

8/ References:

Our original VBMC papers, detailing the algorithm:
- NeurIPS 2018: https://arxiv.org/abs/1810.05558
- NeurIPS 2020: http://arxiv.org/abs/2006.08655

And we have a tl;dr software preprint: https://arxiv.org/abs/2303.09519

7/ PyVBMC is available on both PyPI (pip install pyvbmc) and conda-forge, and provides a user-friendly Pythonic interface.GitHub repo:...

7/ PyVBMC is available on both PyPI (pip install pyvbmc) and conda-forge, and provides a user-friendly Pythonic interface.

GitHub repo: https://github.com/acerbilab/pyvbmc

Detailed notebook tutorials and documentation make it accessible to new and experienced users:
https://acerbilab.github.io/pyvbmc/examples.html

7/ PyVBMC is available on both PyPI (pip install pyvbmc) and conda-forge, and provides a user-friendly Pythonic interface.Docs:...

7/ PyVBMC is available on both PyPI (pip install pyvbmc) and conda-forge, and provides a user-friendly Pythonic interface.

Docs: https://acerbilab.github.io/pyvbmc/

Detailed notebook tutorials make it accessible to new and experienced users:
https://acerbilab.github.io/pyvbmc/examples.html

6/ Like all methods, there are limitations!PyVBMC:- Works up to ~10 continuous model parameters- Works with reasonably smooth target...

6/ Like all methods, there are limitations!

PyVBMC:
- Works up to ~10 continuous model parameters
- Works with reasonably smooth target posteriors (no weird posteriors)
- Can deal with mild multi-modality, but (for now) might not work with very separate modes

5/ PyVBMC uses Gaussian processes and a variational approximation to efficiently compute a flexible (non-Gaussian) approximate posterior...

5/ PyVBMC uses Gaussian processes and a variational approximation to efficiently compute a flexible (non-Gaussian) approximate posterior distribution of the model parameters, from which statistics and posterior samples can be easily extracted.

4/ PyVBMC is particularly effective when the evaluation of the likelihood of your model: - is at least a bit expensive (e.g., ~ one second...

4/ PyVBMC is particularly effective when the evaluation of the likelihood of your model:

- is at least a bit expensive (e.g., ~ one second per evaluation),

OR

- the likelihood evaluation is stochastic (e.g., estimated by Monte Carlo sampling aka simulation).

(or both)

3/ Using PyVBMC is super simple. Define your model (prior, likelihood, a few options), give them to the VBMC optimization process, and go!

3/ Using PyVBMC is super simple.

Define your model (prior, likelihood, a few options), give them to the VBMC optimization process, and go!

2/ PyVBMC works similarly to Bayesian optimization, but with the goal of *inference* (getting the full posterior distribution) instead of...

2/ PyVBMC works similarly to Bayesian optimization, but with the goal of *inference* (getting the full posterior distribution) instead of *optimization* (getting a single point estimate).

Probabilistic numerics FTW.

1/ PyVBMC 1.0 is out! 🎉https://github.com/acerbilab/pyvbmcA new Python package for efficient Bayesian inference.Get a posterior...

1/ PyVBMC 1.0 is out! 🎉

https://github.com/acerbilab/pyvbmc

A new Python package for efficient Bayesian inference.

Get a posterior distribution over model parameters + the model evidence with a small number of likelihood evaluations.

No AGI was created in the process!

I am trying out this new Bayesian optimization algorithm.

I am trying out this new Bayesian optimization algorithm.

Interesting proposal to use inverse binomial sampling for unbiased estimation of the log-likelihood (see https://github.com/acerbilab/ibs)...

Interesting proposal to use inverse binomial sampling for unbiased estimation of the log-likelihood (see https://github.com/acerbilab/ibs) in quantum machine learning: https://arxiv.org/abs/2211.04965

(slightly cross-posted from the bird)

Thorough benchmark showing the superiority of hybrid Bayesian optimization methods, and specifically our BADS...

Thorough benchmark showing the superiority of hybrid Bayesian optimization methods, and specifically our BADS (https://github.com/acerbilab/bads), in several control engineering problems: https://arxiv.org/abs/2211.02571

(cross-posted from the bird)

First toot!New version of the BADS optimization algorithm just released (v1.1.1), with full support for heteroskedastic (user-provided)...

First toot!

New version of the BADS optimization algorithm just released (v1.1.1), with full support for heteroskedastic (user-provided) noise and several fixes: https://github.com/acerbilab/bads

(Also new lab repo! https://github.com/acerbilab)

What about Python? Coming *soon* — for real. Stay tuned!