Nur Lan

Post-doc in cognitive science, ENS Paris

[Philippe Schlenker 🇫🇷 🇪🇺 🇺🇦 👨‍🏫 @phschlenker_ac]

"What the dance language of bees really has in common with human language" - a general audience piece, through CNRS, on our recent article in Biological Reviews.
https://t.co/zmlqP4AoIg

[Nur Lan @nurikolan]

@ArielWaysman תראה smiling friends

[Nur Lan @nurikolan]

למחרת פורסם תיקון

[Nur Lan @nurikolan]

הזדמנות טובה להיזכר בטעות הנפלאה הזו בכותרת הראשית של ״הארץ״ ב-2008.
שר החוץ הצרפתי אמר במקור:
Israel will hit Iran

אלא שבצרפתית אין ה׳ מנושפת, ולכן הרבה צרפתים הוגים hit כמו eat.

התרגום המוזר הצליח לעבור את כל שדרת העריכה של הארץ, וזו היתה התוצאה:

[Nur Lan @nurikolan]

@UriahCanaff

[Matan Abudy @matanabudy]

📄 New paper: "A Minimum Description Length Approach to Regularization in Neural Networks" with Orr Well, Emmanuel Chemla, @roni_katzir, and @nurikolan .
We explore why neural networks often struggle with simple structured tasks.
Spoiler: our regularizers might be the problem.
🧵

[Rɪᴄʜᴀʀᴅ Kᴇᴍᴘ ⋁ @COLRICHARDKEMP]

UK Parliamentary Commission Report on 7 October documents Hamas mass murder, torture, gang rapes & kidnap. Important weapon against the anti-Israel propaganda denying the pure evil of Hamas. Commission chaired by one of UK’s most distinguished historians.

https://t.co/Yu6rJgjxjY

[Nur Lan @nurikolan]

@cedricsueur @bornean_OU Merci, pensez à déménager chez Bluesky :)

[Imry Ziv @imry_ziv]

LLMs are bad theories of human linguistic cognition but can be useful proxies for less bad ones. Here's how and why it might be useful:
https://t.co/OOarnnnRUF

Joint paper with @nurikolan, Emmanuel Chemla, and @roni_katzir

[Nur Lan @nurikolan]

@GaryMarcus @megamor2 Even for RNNs that can potentially count, current training methods prevent them from learning the task:
https://t.co/2YGwJ7B4Bg

[Nur Lan @nurikolan]

⭐️🗞️ Accepted to ACL 2024 main conf! #ACL2024NLP

Neural nets can in theory learn FLs such as aⁿbⁿ & Dyck. Yet no one ever finds such nets using standard techniques. Why?

We suggest that the culprit might have been the objective function all along 👇

https://t.co/JCYrGLZkhs

[Nur Lan @nurikolan]

🗞️🧪 New paper with Emmanuel Chemla and @roni_katzir:

Neural nets offer good approximation but consistently fail to generalize perfectly, even when perfect solutions are proved to exist.

We check whether the culprit might be their training objective.

https://t.co/JCYrGLYMrU
🧵

[Nur Lan @nurikolan]

@aharon1keshales את we're all going to the world's fair ראית?

[Nur Lan @nurikolan]

@damien_leblanc @GrandAction

[Nur Lan @nurikolan]

@matzhikulim https://t.co/UpZpGddqi8

[Nur Lan @nurikolan]

Our findings are in line with works such as El-Naggar et al. (2023) who found similar shortcomings of common objectives for other archs.
https://t.co/SKp5TfsbFD

As well as with our MDL RNNs which achieve perfect generalization on aⁿbⁿ, Dyck-1 and more.
https://t.co/5Jj8CAYqOb

[Nur Lan @nurikolan]

🗞️🧪 New paper with Emmanuel Chemla and @roni_katzir:

Neural nets offer good approximation but consistently fail to generalize perfectly, even when perfect solutions are proved to exist.

We check whether the culprit might be their training objective.

https://t.co/JCYrGLYMrU
🧵

[Nur Lan @nurikolan]

We build an optimal aⁿbⁿ LSTM based on @gail_w et al. (2018) and find that it is not an optimum of the standard cross-entropy loss, even with regularization terms that are expected to lead to good generalization (L1/L2).

Meta-heuristics (early stop, dropout) don't help either.

[Nur Lan @nurikolan]

@AyOdeh כל הכבוד

[Naomi Havron, PhD @NaomiHavron]

Some of these children are my cousin's children.

What can you do? Put pressure on your governments to help free them now!