Fabio De Sousa Ribeiro (@fabayesian) header
Fabio De Sousa Ribeiro (@fabayesian) profile picture

Postdoc at Imperial College London. Working on causality, generative models, and Bayesian inference.

London
Joined July 2023
Fabio De Sousa Ribeiro retweeted
I like em dashes, and I hate that some people now take them as evidence of text being AI generated.
502
495
187
7,096
741,767
Fabio De Sousa Ribeiro retweeted
I like the resurgence of "old-school" NFs that is happening. Here to distill CNFs for free-energy estimation. RL xitter.gq/GhugareRaj/status/1929… Diffusion models (NFDM) xitter.gq/GrigoryBartosh/status/… Latent SDEs (SDE Matching) xitter.gq/canaesseth/status/1887… #SDE #Diffusion #GenAI #ML #FlowMatching
Excited to release FORT, a new regression-based approach for training normalizing flows 🔥! 🔗 Paper available here: arxiv.org/abs/2506.01158 New paper w/ @osclsd @jiarlu @tangjianpku @mmbronstein @Yoshua_Bengio @AlexanderTong7 @bose_joey 🧵1/6
1
8
44
3,731
Why work on medical AI? It's one of the most impactful usecases of AI, it's as simple as that. If you want to work on something meaningful instead of using AI to optimize for clicks/revenue/engagement/etc., consider working on medical AI!
33
33
6
384
38,635
Fabio De Sousa Ribeiro retweeted
We REALLY REALLY need a "Findings" for NeurIPS, ICLR, and ICML. 25,000 submissions at this year's NeurIPS represents extreme excess pressure. It takes valuable time away from legitimate new research. One question is how to administer it. I suggest that Findings go through a lightweight round focused on improving clarity, reeling in overselling, etc. An AC can then "sign off". Authors can always decline the opportunity, if they want to try for the next conference. The NeurIPS's proceedings are called "Advances in Neural Information Processing Systems". We could have "Findings in Neural Information Processing Systems", or to not trample on their brand, perhaps "Contributions to Neural Information Processing Systems". ICLR could have "Letters on Learning Representations." ICML could have "Machinations on Machine Learning."
23
35
3
423
81,630
Fabio De Sousa Ribeiro retweeted
twitter hype is out of control again. we are not gonna deploy AGI next month, nor have we built it. we have some very cool stuff for you but pls chill and cut your expectations 100x!
79
174
12
4,590
305,964
Replying to @liyzhen2
We've got a flavour of this coming out soon 😌
1
497
Fabio De Sousa Ribeiro retweeted
If your learning algorithm is based on correlation rather than causation, it will struggle with overfitting. To understand something is to identify its minimal sufficient causal mechanisms. Parsimony isn't just elegance, it's generalization robustness.
82
181
29
1,675
116,423
Replying to @gabriberton
GANs came out in June 2014 and diffusion models in March 2015, not a huge gap between them
1
241
Fabio De Sousa Ribeiro retweeted
Everything old is new again. Mamba/ssm folks should really google their "new idea + lstm" please. About a decade ago, people have tried a shitton of things with lstms. Nothing wrong with retrying with modern tools, but ack the past. This is not the first such case I see btw.
Mamba2D: A Natively Multi-Dimensional State-Space Model for Vision Tasks paper: arxiv.org/abs/2412.16146 SSMs are emerging as efficient alternatives to transformers but struggle with spatial dependencies in visual tasks due to biases from their natural language processing origins. Existing methods rely on arbitrary 1D scan directions to process 2D data, which limits effectiveness. Mamba2D introduces a native 2D scan direction that models spatial dependencies more effectively by considering both input dimensions simultaneously. On the ImageNet-1K dataset, Mamba2D achieves performance comparable to prior SSM adaptations for vision tasks, offering a more efficient approach to handling visual inputs.
23
58
7
606
120,825
Fabio De Sousa Ribeiro retweeted
Why do NNs often learn similar representations? Existing identifiability results offer theoretical insights, but applying them in practice poses challenges. We’ll present our new work exploring these challenges next week at @unireps #NeurIPS2024 🇨🇦🎉 openreview.net/pdf?id=SQKUZS… 1/
1
9
3
35
3,275
Fabio De Sousa Ribeiro retweeted
A common question nowadays: Which is better, diffusion or flow matching? 🤔 Our answer: They’re two sides of the same coin. We wrote a blog post to show how diffusion models and Gaussian flow matching are equivalent. That’s great: It means you can use them interchangeably.
16
201
28
940
173,509
🎉Thrilled to share our paper "Benchmarking Counterfactual Image Generation" is accepted to @NeurIPSConf 2024!🌟 A big shoutout to my co-first authors Nikos Spyrou, @nefeligk_ and our amazing team: @SnchzPedro_, @vlontzos, Yannis Panagakis, @GiorgosPapanas & @STsaftaris! 🙌
1
3
2
12
1,560
Fabio De Sousa Ribeiro retweeted
I’m thrilled to announce that my paper has been accepted at #UNSURE2024 at #MICCAI2024! Check out the video below where I give a summary of my findings! Keep an eye out next week for a full-length post where I’ll dive deeper into the work and what’s next!
#1: Evaluating Reliability in Medical DNNs by @HarryEJAnthony & @KostasKamnitsas #MICCAI2024 Tl,Dr: New OOD benchmarks reveal artefacts can increase model confidence. A hybrid of feature & confidence-based methods is key! 📜: openreview.net/forum?id=YcS0… 📽️: youtu.be/grKn9m-12j0?feature…
4
24
1,410
Grateful for the warm reception at the #CausalAI Conference ! Honoured to share how #CausalSystemDesign draws parallels between Aristotle’s 4 Causes, NASA TRLs & Pearls causal modelling to transform AI. #AIInnovation #AIforBusiness
2
3
9
676
Replying to @SnchzPedro_
Possibly confounded by the absolute number of publications increasing?
1
27
Fabio De Sousa Ribeiro retweeted
📢 New preprint on finding good representations for Visual Question Answering! We focus in particular on object-centric (OC) learning – since its role in downstream tasks is still partially underexplored – and on pretrained foundation models. 📄 arxiv.org/abs/2407.15589 🧵👇
1
4
1
16
1,831
Fabio De Sousa Ribeiro retweeted
Looking for PhD candidates interested in exploring multi-modal AI for healthcare! If you or someone you know is curious about this opportunity, feel free to reach out. More info about me and the team: cherise215.github.io/
11
130
4
502
56,832