PhD in prob. ML @ APRIL-lab at the university of Edinburgh. Tractable probabilistic ML, closed form stat. quantities and probabilistic ML under hard constraints
Edinburgh/Berlin
Joined December 2009
- Tweets134
- Following216
- Followers53
- Likes9.3K
leander retweeted
Exciting news in the global statistics community! Grace Wahba was awarded the prestigious 2025 International Prize in Statistics for her groundbreaking work on smoothing splines, which revolutionized data analysis and machine learning. https://www. statprize.org/index.cfm
leander retweeted
We learn more expressive mixture models that can subtract probability density by squaring them
🚨We show squaring can reduce expressiveness
To tackle this we build sum of squares circuits🆘
🚀We explain why complex parameters help, and show an expressiveness hierarchy around🆘
leander retweeted
In our ICML 2024 (@icmlconf) paper, we study neurosymbolic methods under the very common independence assumption… and find many problems!
1️⃣ Non-convexity
2️⃣ Disconnected minima
3️⃣ Unable to represent uncertainty
Making optimisation very challenging 😱
Let’s explore 👇
leander retweeted
Will be at AISTATS this week, would love to chat if you have integrals to approximate or are generally into compstat / opt. transport / estimation in causal inference.
Also we have two papers around importance sampling and variational inference.
1st: adaptive IS for heavy tails
leander retweeted
cut holes in your densities to be more expressive!
📜 openreview.net/pdf?id=xIHi5n…
💻 github.com/april-tools/squar…
Classical mixture models are limited to positive weights and this requires learning very large mixtures!
Can we learn (deep) mixtures with negative weights?
Answer in our #ICLR2024 spotlight by @loreloc_ Aleks, Martin, Stefan, Nicolas @arnosolin
📜openreview.net/forum?id=xIHi…
leander retweeted
The softmax bottleneck is an interesting problem; it has many side effects which we do not yet fully understand!
If you want to build an intuition for the problem, here is an interactive visualisation I made grv.unargmaxable.ai/static/f… (best viewed on desktop).
(6/10) This problem is in fact very much related with the softmax bottleneck issue (arxiv.org/abs/1711.03953)
Basically, we try to map "low" dimensional contextual representations to potentially high-dimensional contextual probability manifolds, using a simple linear layer:
leander retweeted
Classical mixture models are limited to positive weights and this requires learning very large mixtures!
Can we learn (deep) mixtures with negative weights?
Answer in our #ICLR2024 spotlight by @loreloc_ Aleks, Martin, Stefan, Nicolas @arnosolin
📜openreview.net/forum?id=xIHi…
After the score-based models, I decided to take a step back and cover the basics in a new blog post: Probabilistic modeling and Mixture Models. Additionally, I make a brief intro to ✨Probabilistic Circuits✨ Check:
📄Post: jmtomczak.github.io/blog/19/…
🖥️ Code: github.com/jmtomczak/intro_d…
Great opening of the ProbNum School from @PhilippHennig5 and Ilse Ipsen. I wonder though: Is everything we ever do linear algebra and piecewise-non linearities?
