Jörg Schlötterer retweeted
Do you want to know what information LLMs prioritize in text synthesis tasks? Here's a short 🧵 about our new paper: an interpretable framework for salience analysis in LLMs.
First of all, information salience is a fuzzy concept. So how can we even measure it?
Jörg Schlötterer retweeted
🎉2 papers accepted to #NAACL2025 Main! All about potential malicious uses of model editing. If a method can effectively modify a model, with anyone able to upload whatever models to Hugging Face or GitHub, we @paul_y0u @SeifertChristin @jschloetterer feel it’s a bit scary😅
Jörg Schlötterer retweeted
Especially happy about the first PhD paper of @funzac83 - she's attending in person, so say hello to her! (1) is her joint work w @belaneuendorf @JoanPlepi Jörg Schlötterer and @ytaipsw
Jörg Schlötterer retweeted
'The Queen of England is not England’s Queen.' Even though neither is true anymore, we would expect that if a fact is "known", the reverse fact is also "known". Turns out, this assumption does not hold for PLMs.
aclanthology.org/2024.findin…
@paul_y0u
@jschloetterer
Jörg Schlötterer retweeted
When we (or LLMs) explain technical texts, those explanations can be vague or omit detail. How can we characterize such information loss and help lay readers intuitively recover it?
Excited to share InfoLossQA! 🧵
Paper: arxiv.org/abs/2401.16475
Website: InfoLossQA.ikim.nrw/
Jörg Schlötterer retweeted
Paper is out! Together with Camelia Oprea and @mstrohm, we’ve explored the potential impact of toxic comments on Wikipedia
academic.oup.com/pnasnexus/a… – at Sydney, New South Wales
Jörg Schlötterer retweeted
🚨🚨 The first State of Crypto report "The socio-political landscape of the community" is out! 170+ pages in which we report the results of our first survey including respondents from 170+ countries :)🧵
Jörg Schlötterer retweeted
Grateful to have been awarded Best Parallel Talk at #IC2S2’23! Thanks for the opportunity to present our work, it’s been a pleasure. If you want to find out more about how we add interpretability to pre-trained word embeddings, check out our paper➡️bit.ly/473xfcs
Jörg Schlötterer retweeted
GRAN is superior to GraphRNN:
we investigated on evaluation methods for graph generative models in arxiv.org/abs/2307.06709
Generating graphs similar to samples observed from a dataset has various interesting applications:
Jörg Schlötterer retweeted
Happy to share this recent preprint of our survey on similarity measures for neural network models, which I have worked on with Max Klabunde, @mstrohm, and Florian Lemmerich:
arxiv.org/abs/2305.06329
Jörg Schlötterer retweeted
‼️ 1st #ThirdParty Call open!
Help us building an #OpenWebIndex.
We call for proposals and invite you, researchers + innovators, to submit your applications and join our mission. Get the info package ➡️ openwebsearch.eu/call1... #funding #opencall
Jörg Schlötterer retweeted
❗1st #ThirdParty #OpenCall coming up
Join OpenWebSearch.EU and get involved!
➡️ Opens: March 1 2023
➡️ Track 1: legal and economic aspects of #OpenSearch
➡️ Track 2: legally compliant data acquisition considering societal constraints
More info: openwebsearch.eu/third-party…
Jörg Schlötterer retweeted
Turns out that it's surprisingly easy to bypass ChatGPT's filters.
Discussing an imaginary world populated by human-like animals reveals inherent biases that #ChatGPT has learned.
tw racism
Jörg Schlötterer retweeted
Explainable AI methods help us understand what machine learning black boxes do. But what can we say about the quality of the explanation methods?
Survey: dl.acm.org/doi/10.1145/35835…
Interactive Website: utwente-dmb.github.io/xai-pa…
#xai #evaluation
Jörg Schlötterer retweeted
Have you submitted to IDA yet? Here's more info on why you should submit to IDA: ida2023.org/why-ida/
Jörg Schlötterer retweeted
Meike Nauta is winner of the Zeiss Women Award 2022.
Congratulations Meike!
Congrats Meike Nauta for winning the Zeiss Women Award! #proudsupervisor @IkimUme @unidue @UTwente @UniklinikEssen @ZEISS_Group
Jörg Schlötterer retweeted
Excited to share recent works and experience in #AI+ #health at the University of Duisburg-Essen. Thank you, Christin. Looking forward to it!
@SeifertChristin @unidue
#deeplearning #timeseries #ubiquitous
Jörg Schlötterer retweeted
Happy to introduce our second keynote speaker at the MICCAI MAD workshop: Sandy Engelhardt
@SandyEngelh
🥳#MICCAI2022 @yash_j_sharma
@STsaftaris @MichaelKamp7 @DanielRueckert @KaiUeltzhoeffer @ShadiAlbarqouni @medai_ahmadi
@cmseibold @anirbanakash @affrangi @maierhein
Jörg Schlötterer retweeted
The exact date for the MICCAI MAD workshop is fixed as 22.09.2022 🌼 #MICCAI2022
Check out our website: mad.ikim.nrw