[1/4] Introducing “A Primer on the Inner Workings of Transformer-based Language Models”, a comprehensive survey on interpretability methods and the findings into the functioning of language models they have led to.

ArXiv: arxiv.org/pdf/2405.00208

I'm super happy to co-organize the (Mechanistic) Interpretability social at #ICLR2024 with Nikhil Prakash 🛩️ ICLR2024! 🔍

If you plan to attend, help us make this meetup awesome by filling the form below! 😄

cc Explainable AI Neel Nanda EleutherAI Anthropic Cas (Stephen Casper)

Hiring a Post-doctoral researcher in my lab, please consider applying if you see the position fits your expertise and interest. The research is at the intersection of NLP and CV leveraging neuro-symbolic modeling in V&L navigation; #NLProc , #ACL , #CVPR careers.msu.edu/en-us/job/5182… .

On April 20th, I’ll run for Spierfonds to raise money for research on muscular diseases. Your donation here would greatly please me :)
spieractie.nl/fundraisers/ho…

Circuits are a hot topic in interpretability, but how do you find a circuit and guarantee it reflects how your model works?

We (Sandro Pezzelle, Yonatan Belinkov, and I) introduce a new circuit-finding method, EAP-IG, and show it finds more faithful circuits arxiv.org/abs/2403.17806 1/8

Still looking at self-attention weights to see how (token) representations are being mixed in a Transformer layer? Try more informative alternatives by simply enabling ‘output_context_mixings’ in your HuggingFace model call:
Notebook: colab.research.google.com/drive/114Yigbe…

#NLProc

The moment for “Transformer-specific Interpretability” tutorial is very approaching!
Check out our detailed program & an extensive suggested reading list here: projects.illc.uva.nl/indeep/tutoria…

🗓️ Thursday, 21 March
⏱️ 14:00-17:30 pm
📍 Radisson room
#EACL2024 #NLProc #XAI

Today I’ll present my joint work with Christof Monz on “Analyzing the Evaluation of Cross-lingual Knowledge Transfer in Multilingual Language Models” at #EACL2024
Happy to see everyone there!
📍 Marie Louise 1, Multilingual session
🕟 4 p.m.
📖 arxiv.org/abs/2402.02099

Sharing some great news today. My paper titled 'Encoding of lexical tone in self-supervised models of spoken language' has been accepted to the main NAACL 2024 conference!
Many thanks to my co-authors Michaela Watkins afra alishahi Arianna Bisazza Grzegorz Chrupała 🇪🇺🇺🇦

A brief 🧵below:

I really enjoyed the visit and chats with lively CardiffNLP folks, special thanks to Taher Pilehvar, Jose Camacho-Collados, and kiamehr for having me!🤗

Wow!! This is great news🥳
Thanks Inseq! I’m excited to try it out!

We work on human-centered natural language processing and I am hiring a PhD candidate and a PostDoc. Please let me know if you are interested or share with someone who might be.

PostDoc description: uni-goettingen.de/en/644546.html…
PhD description: uni-goettingen.de/en/644546.html…

Over the coming two weeks, I’ll have a great opportunity to share my research through a series of talks at:
- CSTR, Edinburgh (Feb 26)
- CardiffNLP (Feb 29)
- TeIAS, Tehran (Mar 6)
- InDeep workshop, Amsterdam (Mar 8)
Excited to meet, discuss & learn from everyone there!

Arthur’s Seat view from Informatics’ terrace serves as a refresh button during my break times!
Very enjoying my visit w/ Ivan Titov and his team at EdinburghNLP!

Happy to share our paper on LM’s Feed-forward network (FF) analysis has been accepted as an #ICLR2024 spotlight!
💡FF boosts attention between words forming compound nouns, named entities, etc.
💡FF and LayerNorm cancel out each other’s effects

📄arxiv.org/abs/2302.00456

For people interested in #CausalNLP and Interpretability & Robustness of #LLMs , I just shared my Keynote Talk at #EMNLP2023 BlackboxNLP workshop: 'Causal NLP: A Path towards Opening the Black Box of NLP'
- 🎦Talk video: youtube.com/watch?v=Auls8a…
- 💻Slides: docs.google.com/presentation/d…

Even more grateful to also receive an Outstanding Paper Award at #EMNLP2023 for our paper on cross-lingual consistency of factual knowledge in LMs! with Jirui Qi and Raquel Fernández.
@GroNLP AmsterdamNLP

arxiv.org/abs/2310.10378

Fantastic news! Outstanding paper award for this paper on opening up the blackbox of neural speech recognition models at #EMNLP2023 !

by Hosein Mohebbi (Tilburg & InDeep) and collaborators from Tilburg and Amsterdam.

projects.illc.uva.nl/indeep/