André Martins

So nice! Where did you get it?

Applications for the 2025 Lisbon Machine Learning Summer School (LxMLS) are open, with @andre-t-martins.bsky.social as one of the organizers.
LxMLS is a great opportunity to learn from top speakers and to interact with other students. You can apply for a scholarship.

Apply here:
lxmls.it.pt/2025/

Visit our poster soon at #AISTATS2025! 🚀

@istecnico.bsky.social @itnewspt.bsky.social @ellisunitlisbon.bsky.social

Made in @sardine-lab-it.bsky.social !

Our experiments show competitive / superior results in terms of coverage, efficiency, and adaptiveness compared to standard non-conformity scores based on softmax, such as InvProb and RAPS. 7/N

As a bonus, for softmax (which is not sparse) we also obtain a new “log-margin” non-conformity score which is the log-odds ratio between the most probable class and the true one. 6/N

For γ-entmax (which recovers sparsemax with γ=2 and softmax with γ=1), the non-conformity scores use the L_δ-norm (with δ = 1 / (γ - 1)) instead of the L_1 norm. 5/N

The answer is yes! For sparsemax, this corresponds to a new non-conformity score which accumulates the absolute differences of logits up to the target label. 4/N

These sparse transformations have a temperature parameter which controls the amount of sparsity. Can we use split conformal prediction to calibrate this temperature parameter and return sparse sets with coverage guarantees? 3/N

Conformal prediction quantifies uncertainty by predicting *sets* instead of points, offering coverage guarantees. Sparse transformations (sparsemax, entmax, etc.) are softmax alternatives which return sparse probability vectors, useful to select a subset of relevant labels. 2/N

Our upcoming #AISTATS2025 paper is out: “Sparse Activations as Conformal Predictors” with Margarida Campos, João Calém, Sophia Sklaviadis, and
@marfig.bsky.social:

arxiv.org/abs/2502.14773.

This paper puts together two lines of research, dynamic sparsity and conformal prediction. 🧵

@andre-t-martins.bsky.social talking about multilingual LLMs at the @priberam.bsky.social Machine Learning Lunch Seminar in @istecnico.bsky.social,
@ellisunitlisbon.bsky.social

$\infty$-Video: A Training-Free Approach to Long Video Understanding via Continuous-Time Memory Consolidation Current video-language models struggle with long-video understanding due to limited context lengths and reliance on sparse frame subsampling, often leading to information loss. This paper introduces $...

🎉 New paper by Saul Santos in collaboration with @tozefarinhas.bsky.social and @andre-t-martins.bsky.social!! 🎉

∞-Video: A Training-Free Approach to Long Video Understanding via Continuous-Time Memory Consolidation

Paper: arxiv.org/abs/2501.19098