Matteo Pinna

Incoming ELLIS PhD @ Hochreiter’s lab | Formerly CIML @Unipisa

[Matteo Pinna @nennomp]

ParalESN made it to ICML 🚀

We revisit the reservoir computing paradigm to enable parallel recurrence and higher-dimensional reservoirs.

Preprint: https://t.co/UsvaASiYie
Co-authors: G. Lagomarsini, A. Ceni, @claudiogallicc1

#ICML #ICML2026

[Matteo Pinna @nennomp]

@matteo_achilli Non sapevo che chiedere tre colori fosse una domanda a trabocchetto. Chiedo venia.

[ELLIS @ELLISforEurope]

🎓 We're thrilled to welcome 121 new PhD students to the ELLIS network from our 2025/2026 call — selected from a record-breaking 2,100+ submitted applications, a 43% increase from last year.

👉 Read the full article for more details: https://t.co/II3rkmOIOm

[Matteo Pinna @nennomp]

@Guy_T_Sky @furongh Right, wasn't thinking about that.

[Matteo Pinna @nennomp]

@Guy_T_Sky @furongh Have the same question. The approach is also more (memory-) efficient so, unless I missed something, it cannot be due to compute, no?

[Matteo Pinna @nennomp]

@TaimingLu Cool insights.

Do you have any intuition on why the very-strong teacher setting underperforms with respect to the strong teacher one?

[Andrew Gordon Wilson @andrewgwils]

This is your annual reminder that we don’t need to speculate about whether we will have a “theory of deep learning” and what form it might take, because we already have a basic understanding of generalization in deep learning: https://t.co/AgHdSQjCvU

[Matteo Pinna @nennomp]

Great initiative.

Hopefully this becomes the norm and many more conferences adopt something similar.

[ICML Conference @icmlconf]

Reviewers & ACs for #ICML2026 have been recognized for their service!

- Reviewers: 4439 Gold (free registration), 4437 Silver. 17749 total reviewers were assigned >= 1 paper
- ACs: 1647 receive free registration, out of 1691 who were assigned >= 1 paper

TY for your hard work!

[Matteo Pinna @nennomp]

@DJiafei Besides, most of the time, AI generated figures look very ugly, no?

Also, the first figure of the paper is hands down the most important one.

[Konstantin Dobler @konstantdobler]

A plot of ICLR papers by country is making the rounds, showing no EU + Japan papers and people are drawing all kinds of conclusions.

..but the plot excludes all (!) EU institutions due to a cutoff. China + USA still dominant of course but the full picture looks a bit different.

[Claudio Gallicchio @claudiogallicc1]

🚀 Reservoir Computing is officially back in the Deep Learning game!

Thrilled to announce our paper "ParalESN" has been accepted at #ICML2026! 🎉

We resurrected RC for the DL era, finally unlocking parallel processing via associative scan.

Preprint: https://t.co/dYPBkCpGeq🧵👇

[Matteo Pinna @nennomp]

@mhrezaeics AI-assisted research does not only generate slop, it also makes good labs more productive (i.e., higher number of good submissions).

I would also wait a couple more cycles before judging the acceptance rate trajectory.

[Matteo Pinna @nennomp]

@jgh0sh @ryancjulian Maybe patents? Or, if the idea is exceptional, you go for the startup route. I guess OSS is another option as well.

[Matteo Pinna @nennomp]

@jgh0sh @ryancjulian Imho, good taste means you get published and cited a lot, while personal preference doesn't guarantee the above.

[Matteo Pinna @nennomp]

@shaohua0116 In some cases there may be some fundamental limitation in the paper that does not warrant a higher score. Fine with that, but I would prefer them being explicit about it or tell me exactly what needs to be addressed in order for them to adjust the score.

[William Merrill @lambdaviking]

[1/8] New paper with Hongjian Jiang, @YanhongLi2062, Anthony Lin, @Ashish_S_AI:

📜Why Are Linear RNNs More Parallelizable?

We identify expressivity differences between linear/nonlinear RNNs and, conversely, barriers to parallelizing nonlinear RNNs 🧵👇

[Matteo Pinna @nennomp]

@gabriberton Agree.

I argue that a good percentage of those complaining are the ones affected by the desk rejections. Otherwise it's absurd to complain about mechanisms introduced to protect authors from such low-effort reviews...

[Mayank Mishra @MayankMish98]

Introducing M²RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling

We bring back non-linear recurrence to language modeling and show it's been held back by small state sizes, not by non-linearity itself.

📄 Paper: https://t.co/AS8e2tNrRa
💻 Code: https://t.co/LMvBcI22Du
🤗 Models: https://t.co/NCmjrpNriq

[Shuangfei Zhai @zhaisf]

Fun fact: In an N layer Transformer, you can drop the QK projections in the top N/2 layers and reuse the attention from layer N/2, and it basically works.

The top half layers now act like an MLP mixer with its mixing weights (attention) generated by a Transformer (the bottom half).