iiinova — where your digital self keeps evolving

Machine Learning in Science @mackelab.bsky.social · 06/07/2026

ModelSMC: we frame LLM-based scientific model discovery as Bayesian inference. Sequential Monte Carlo over executable model structures, with the LLM as a probabilistic proposal mechanism. 📍 Wed Jul 8, at 2:30–4:15 PM KST in Hall A #3512 📄 arxiv.org/abs/2602.18266 🧵 bsky.app/profile/mack...

0113

Machine Learning in Science @mackelab.bsky.social · 06/07/2026

PRISM: Model inference at scale. We infer joint posteriors over model structure & parameters, with test-time control of complexity via a tunable prior. 📄 arxiv.org/abs/2603.15292 by Manuel Gloeckler, @josepemanzano.bsky.social, @mristam.bsky.social, @coschroeder.bsky.social, @jakhmack.bsky.social

171

Machine Learning in Science @mackelab.bsky.social · 02/07/2026

We congratulate @matthijspals.bsky.social on his successful PhD defense! During his time in the Mackelab, he used RNNs to link neural activity with underlying mechanisms. Now he moved on to a Postdoc position in @durstewitzlab.bsky.social at the ZI Mannheim and the University of Heidelberg.

Left to right: Philipp Berens, Anna Levina, Matthijs Pals, Peter Dayan, Jakob Macke.

2374

Machine Learning in Science @mackelab.bsky.social · 26/06/2026

A unique benefit of our framework? We can analyse the model posterior! We semantically cluster the generated HH models by their added ion-channel mechanisms. Low-weight clusters flag structural mismatch scientists can discard; high-weight clusters point to the mechanisms actually worth pursuing.

110

Machine Learning in Science @mackelab.bsky.social · 26/06/2026

We treat Hodgkin-Huxley models as a structural problem: infer missing ion-channel mechanism. ModelSMC uncovers mechanistically meaningful extensions, exposing structural ambiguities and improving fit, yielding a posterior over plausible models rather than a single best model.

110

Machine Learning in Science @mackelab.bsky.social · 26/06/2026

In a pharmacological kidney model, ModelSMC recovers meaningful structure from real-data. Even with highly limited data, the posterior concentrates on a small class of aldosterone feedback laws.

110

Machine Learning in Science @mackelab.bsky.social · 26/06/2026

As one concrete instantiation of this framework, we introduce ModelSMC: Sequential Monte Carlo directly in model space. We keep a population of candidate programs as particles, propose refinements via the LLM, weight by likelihood, and resample.

110

Machine Learning in Science @mackelab.bsky.social · 26/06/2026

Most LLM discovery systems are agentic loops: propose, evaluate, keep the best one. This gives a single point estimate, discarding the probabilistic structure scientists rely on to know what to rule out, what is uncertain, and where more data would help. Instead, we frame discovery as inference.

120

Machine Learning in Science @mackelab.bsky.social · 26/06/2026

New paper: We recast automated scientific model discovery with LLMs as Bayesian inference! LLMs write code and carry domain knowledge, great for proposing models. The key idea: discovery is inference, not just generation. What distribution of models explains the data? 🧵 arxiv.org/abs/2602.18266

1124

Machine Learning in Science @mackelab.bsky.social · 30/03/2026

Come and work with us! We have a PostDoc position at the intersection of ML and Biogeoscience within the TERRA excellence cluster @terra-cluster.org, w/ Senckenberg. Be part of a great ML and Geo community and use ML to investigate fire and its impact on global vegetation🔥 🌱🌳 www.mackelab.org/jobs/

074

Machine Learning in Science @mackelab.bsky.social · 11/03/2026

Friday, 13:15 (Poster 2-040): @lulmer.bsky.social presents “Neural activity constraints improve task-optimized connectome-constrained models” (joint work with @srinituraga.bsky.social).

120

Machine Learning in Science @mackelab.bsky.social · 16/02/2026

Auguste developed ML methods for linking neural activity and behavior, ranging from classic discriminative models to deep generative models such as VAEs and DDPMs e.g. doi.org/10.52202/079... or doi.org/10.1016/j.ce....

0130

Machine Learning in Science @mackelab.bsky.social · 16/02/2026

Janne developed deep mechanistic networks to study when and how detailed measurements of brain wiring (connectomes) can enable accurate, neuron-level predictions of neural dynamics across the brain (www.nature.com/articles/s41...).

1100

Machine Learning in Science @mackelab.bsky.social · 13/01/2026

Guy Moss (@gmoss13.bsky.social) developed and applied simulation-based inference methods to solve inference problems in glaciology, in collaboration with @geophys-tuebingen.bsky.social. E.g., openreview.net/forum?id=yB5... 3/3

180

Machine Learning in Science @mackelab.bsky.social · 13/01/2026

Julius Vetter (@vetterj.bsky.social) worked on deep generative modeling and simulation-based Bayesian inference, with applications to (physiological) time series data. E.g., openreview.net/forum?id=kN0... 2/3

280

Machine Learning in Science @mackelab.bsky.social · 13/01/2026

Happy 2026 everyone! Two freshly minted PhDs 🧑‍🎓emerged from our lab at the end of last year. We congratulate Dr Julius Vetter (@vetterj.bsky.social) and Dr Guy Moss (@gmoss13.bsky.social)! Here seen celebrating with the lab 🎳. 1/3

1183

Machine Learning in Science @mackelab.bsky.social · 01/12/2025

Fifth, we bring AutoML to SBI pipelines with a practical performance metric that does not require ground-truth posteriors, improving inference quality on the SBI benchmark! By @swagatam.bsky.social, @gmoss13.bsky.social, @keggensperger.bsky.social, @jakhmack.bsky.social 10/11

140

Machine Learning in Science @mackelab.bsky.social · 01/12/2025

Fourth, in collaboration with Ian C Tanoh and Scott Linderman, we used the Jaxley toolbox and extended Kalman filters to estimate the marginal log-likelihood of a biophysical neuron model. We showed that this enables identifying biophysical parameters given extracellular recordings. 8/11

140

Machine Learning in Science @mackelab.bsky.social · 01/12/2025

Third, in collaboration with @kyrakadhim.bsky.social, @philipp.hertie.ai, and others, we built a task- and data-constrained biophysical network of the outer plexiform layer of the mouse retina. To optimize this model, we built it on top of our Jaxley toolbox for differentiable simulation. 6/11

131

Machine Learning in Science @mackelab.bsky.social · 01/12/2025

Second, come by to check out NPE-PFN: We leverage the power of tabular foundation models for training-free and simulation-efficient SBI. SBI has never been so effortless! By @vetterj.bsky.social, Manuel Gloeckler, @danielged.bsky.social, @jakhmack.bsky.social 4/11

120

Machine Learning in Science @mackelab.bsky.social · 01/12/2025

First, we introduce FNOPE, a new simulation-based inference approach for efficiently and flexibly inferring function-valued parameters. By @gmoss13.bsky.social, @leahsmuhle.bsky.social, Reinhard Drews, @jakhmack.bsky.social and @coschroeder.bsky.social 2/11

150

Machine Learning in Science @mackelab.bsky.social · 21/11/2025

We provide three examples and code to demonstrate the complete workflow: gravitational wave parameter estimation (astrophysics), psychophysical model fitting (cognitive science), and ion channel inference (neuroscience).

100

Machine Learning in Science @mackelab.bsky.social · 21/11/2025

We present a structured workflow with practical guidelines for each step: simulator setup, prior specification, method selection, network training, and validation. We illustrate each stage with concrete implementation details and common pitfalls.

110

Machine Learning in Science @mackelab.bsky.social · 02/10/2025

Congrats to Dr Michael Deistler @deismic.bsky.social, who defended his PhD! Michael worked on "Machine Learning for Inference in Biophysical Neuroscience Simulations", focusing on simulation-based inference and differentiable simulation. We wish him all the best for the next chapter! 👏🎓

0301

Machine Learning in Science @mackelab.bsky.social · 30/09/2025

The Macke lab is well-represented at the @bernsteinneuro.bsky.social conference in Frankfurt this year! We have lots of exciting new work to present with 7 posters (details👇) 1/9

1309

Machine Learning in Science @mackelab.bsky.social · 23/07/2025

But does it scale to complex real-world problems? We tested it on two challenging Hodgkin-Huxley-type models: 🧠 single-compartment neuron 🦀 31-parameter crab pyloric network NPE-PF delivers tight posteriors & accurate predictions with far fewer simulations than previous methods.

Results on the pyloric simulator.
(a) Voltage traces from the experimental measurement (top) and a posterior predictive simulated using the posterior mean from TSNPE-PF as the parameter (bottom).
(b) Average distance (energy scoring rule) to observation and percentage of valid simulation from posterior samples; compared to experimental results obtained in Glaser et al.

110

Machine Learning in Science @mackelab.bsky.social · 23/07/2025

What you get with NPE-PF: 🚫 No need to train inference nets or tune hyperparameters. 🌟 Competitive or superior performance vs. standard SBI methods. 🚀 Especially strong performance for smaller simulation budgets. 🔄 Filtering to handle large datasets + support for sequential inference.

SBI benchmark results for amortized and sequential NPE-PF.
C2ST for NPE, NLE, and NPE-PF across ten reference posteriors (lower is better); dots indicate averages and bars show 95% confidence intervals over five independent runs.

100

Machine Learning in Science @mackelab.bsky.social · 23/07/2025

The key idea: TabPFN, originally trained for tabular regression and classification, can estimate posteriors by autoregressively modeling one parameter dimension after the other. It’s remarkably effective, even though TabPFN was not designed for SBI.

Illustration of NPE and NPE-PF: Both approaches use simulations sampled from the prior and simulator. In (standard) NPE, a neural density estimator is trained to obtain the posterior. In NPE-PF, the posterior is evaluated by autoregressively passing the simulation dataset and observations to TabPFN.

110

Machine Learning in Science @mackelab.bsky.social · 11/06/2025

We obtain posterior distributions over ice accumulation and melting rates for Ekström Ice Shelf over the past hundreds of years. This allows us to make quantitative statements about the history of the atmospheric and oceanic conditions.

121

Machine Learning in Science @mackelab.bsky.social · 11/06/2025

Thanks to great data collection efforts from @geophys-tuebingen.bsky.social and @awi.de, we can apply this approach to Ekström Ice Shelf, Antarctica.

100

Machine Learning in Science @mackelab.bsky.social · 11/06/2025

We develop a simulation-based-inference workflow for inferring the accumulation and melting rates from measurements of the internal layers.

100

Machine Learning in Science @mackelab.bsky.social · 11/06/2025

Radar measurements have long been used for measuring the internal layer structure of Antarctic ice shelves. This structure contains information about the history of the ice shelf. This includes the past rate of snow accumulation at the surface, as well as ice melting at the base.

100

Machine Learning in Science @mackelab.bsky.social · 25/04/2025

In FNSE, we only have to solve a smaller and easier inverse problem; it does scale relatively easily to high-dimensional simulators. We validate this on a high-dimensional Kolmogorov flow simulator with around one million data dimensions.

110

Machine Learning in Science @mackelab.bsky.social · 25/04/2025

We apply this approach to various SBI methods (e.g. FNLE/FNRE), focusing on FNSE. Compared to NPE with embedding nets, it’s more simulation-efficient and accurate across time series of varying lengths.

100

Machine Learning in Science @mackelab.bsky.social · 25/04/2025

We propose an SBI approach that can exploit Markovian simulators by locally identifying parameters consistent with individual state transitions. We then compose these local results to obtain a posterior over parameters that align with the entire time series observation.

100

Machine Learning in Science @mackelab.bsky.social · 27/03/2025

Does any of this sound interesting to you? You might be excited to know that we are hiring PhD students and Postdocs! Go to www.mackelab.org/jobs/ for details on more projects and how to contact us, or just find one of us (Auguste, Matthijs, Richard, Zina) at the conference—we're happy to chat!

010

Machine Learning in Science @mackelab.bsky.social · 27/03/2025

Finally, @rdgao.bsky.social is in another Tuesday workshop: “What biological details matter at mesoscopic scales?” He will tell you about the good, the bad, and the ugly of getting more than what you had asked for using mechanistic models and probabilistic machine learning. 17:00, Corriveau/Sateux

121

Machine Learning in Science @mackelab.bsky.social · 27/03/2025

We're also at the workshops! Tuesday at 11:20, @auschulz.bsky.social will give a talk about deep generative models - VAEs and DDPMs - for linking neural activity and behavior, at the workshop on "Building a foundation model for the brain" (Soutana 1).

100

Machine Learning in Science @mackelab.bsky.social · 27/03/2025

Interested in obtained models that are both interpretable and that can generate realistic neural data? @matthijspals.bsky.social will give a contributed talk on fitting stochastic low-rank RNNs to neural data, Saturday 10:00 at the main conference! 🤩

120

Machine Learning in Science @mackelab.bsky.social · 27/03/2025

In the same Friday session (Poster 2-039), @rdgao.bsky.social presents AutoMIND: using machine learning to infer parameters of spiking neural network models from brain recordings. Come by if you hate fitting spiking networks or love degeneracy and invariances! www.biorxiv.org/content/10.1...

120

Machine Learning in Science @mackelab.bsky.social · 27/03/2025

On Friday at 13:15 (Poster 2-012), Zina will talk about the role of recurrent connectivity for motion detection in the fruit fly. Using connectome-constrained models of its visual system, we show how ablating recurrence decreases motion selectivity.

100

Machine Learning in Science @mackelab.bsky.social · 27/03/2025

On Thursday at 20:00 (Poster 1-108), @auschulz.bsky.social introduces LDNS, a diffusion-based latent variable model to generate diverse neural spiking data flexibly conditioned on external variables.

100

Machine Learning in Science @mackelab.bsky.social · 27/03/2025

The @mackelab.bsky.social is represented at @cosynemeeting.bsky.social #cosyne2025 in Montreal with 3 posters, 2 workshop talks, and a main conference contributed talk (for the very first time in Mackelab history 🎉)!

1164

Machine Learning in Science @mackelab.bsky.social · 09/12/2024

We introduce a new maximum-entropy, sample-based approach to solve the source distribution estimation problem. Poster #4006 (East; Fri 13 Dec 11PT). By @vetterj.bsky.social, @gmoss13.bsky.social ➡️ openreview.net/forum?id=0cg... 4/4

0101

Machine Learning in Science @mackelab.bsky.social · 09/12/2024

@jkapoor.bsky.social and @auschulz.bsky.social introduce LDNS, a diffusion-based latent variable model to generate diverse neural spiking data flexibly conditioned on external variables Poster #4010 (East; Wed 11 Dec 11PT) ➡️ openreview.net/forum?id=ZX6... 3/4

1112

Machine Learning in Science @mackelab.bsky.social · 09/12/2024

We show how to generate long sequences of realistic neural data with stochastic low-rank RNNs - and how to find their fixed points. Poster #3908 (East; Wed 11 Dec 11PT). @matthijspals.bsky.social, @aesagtekin.bsky.social, F Pei, M Gloeckler, @jakhmack.bsky.social ➡️ openreview.net/forum?id=C0E... 2/4

1100

Machine Learning in Science @mackelab.bsky.social · 14/11/2024

Hi world! This is the brand-new BlueSky account of the Machine Learning in Science (@jakhmack.bsky.social) lab. We create probabilistic #MachineLearning and #AI Tools for scientific discovery — but more on that soon! For now let’s introduce ourselves with some pictures of our recent group retreat.

0195