Data Science @ Regensburg goes Berlin: iSchool Special

Dear all,

Need a reason to plan a trip to Berlin? The wait is over. It is a pleasure to announce our Data Science @ Regensburg Meetup Roadshow. Apart from the location the setup is as always: a relaxed forum to discuss data science issues with a broad mix of attendees from industry, academia and beyond. What's more, this event also represents the first activity we run as part of our European iSchool Industry agenda (https://www.ischools.org/industry-partnerships) .

And now we already have two confirmed speakers: Alexander Buchholz (Senior Applied Scientist at Amazon) who will talk about music recommendation. And here is our second speaker: Ansar Aynetdinov (PhD student in NLP at Humboldt University Berlin) who will talk about OpinionGPT. There is a common theme underlying both talks ... biases. In one case you want to avoid them and in the second case you want to make these biases visible. I am looking forward to a lively discussion!

More details on both talks below ...

See you in Berlin?
Udo

Speaker
Alexander Buchholz (Senior Applied Scientist at Amazon).

Abstract
Amazon Music ranks millions of items to personalize its customer experience. In this talk we will provide a perspective on the scientific questions that we encounter for training and evaluating learning-to-rank models. We will touch upon topics like unbiased off-policy learning and evaluation, data collection and model deployment from an industry point of view. We will highlight our latest contributions and discuss the challenges that we face on a daily basis.

Short Bio
Alexander Buchholz is a senior applied scientist at Amazon Music in Berlin. Prior to joining Amazon Music he was a postdoc at the University of Cambridge, UK. He received his PhD in computational statistics from the University Paris Saclay and studied Economics, Mathematics and Statistics in Berlin, Paris and Cambridge, MA.

----------------

Speaker
Ansar Aynetdinov (PhD student in NLP at Humboldt University Berlin)

Title
OpinionGPT: Modelling Explicit Biases in Instruction-Tuned LLMs

Abstract
Human biases inevitably get picked up by Large Language Models (LLMs). Consequently, a considerable amount of AI research is dedicated to bias mitigation, since certain societal biases are undesirable in LLM-based end products. In this talk Ansar will talk about OpinionGPT - an instruction-tuned LLM that flips the objective and explicitly models various human biases. He will showcase how different types of biases in the training data affect the way OpinionGPT responds to questions. He will also detail the training process together with associated challenges behind OpinionGPT, as well as discuss general implications of training LLMs on data reflecting specific societal biases.

Short Bio
Ansar is a PhD student at the Machine Learning Chair at the Humboldt University under the supervision of Prof. Alan Akbik. His main research interests are training LLMs in a data- and parameter-efficient manner, as well as LLM evaluation.