BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Microsoft AI&amp\;Pizza event- 13 June 2024 - Speaker to be confir
 med
DTSTART:20240613T163000Z
DTEND:20240613T180000Z
UID:TALK217840@talks.cam.ac.uk
CONTACT:Kimberly Cole
DESCRIPTION:Dear Cambridge AI & Machine Learning Enthusiast\,\n \nThe AI &
  Pizza talk series rolls on! We're excited to announce our next AI & Pizza
  talks event\, scheduled for at 5:30 pm on Thursday\, June 13th\, 2024. Sa
 ve your date for the latest advancements in AI and machine learning\, righ
 t here in Cambridge.\n \nP.S.\, We are constantly looking for speakers fro
 m related fields. Please do reach out to us (chaoma@microsoft.com\; wenbog
 ong@microsoft.com ) if you are interested!\n \nLocation: The Small Lecture
  Theatre\, 21 Station Rd\nDate: Thursday\, June 13th\, 2024\nAgenda:\n5:30
  pm – 6:00 pm: Talks\n6:00 pm – 7:00 pm: Networking\, Pizza\, and Refr
 eshments\n\n\nSpeaker: Meyer Scetbon (Microsoft Research)\nTitle: FiP: a F
 ixed-Point Approach for Causal Generative Modeling\n5:30 pm - 5:45 pm\nAbs
 tract:  Modeling true world data-generating processes lies at the heart of
  empirical science. Structural Causal Models (SCMs) and their associated D
 irected Acyclic Graphs (DAGs) provide an increasingly popular answer to su
 ch problems by defining the causal generative process that transforms rand
 om noise into observations. However\, learning them from observational dat
 a poses an ill-posed and NP-hard inverse problem in general. In this work\
 , we propose a new and equivalent formalism that does not require DAGs to 
 describe them\, viewed as fixed-point problems on the causally ordered var
 iables\, and we show three important cases where they can be uniquely reco
 vered given the topological ordering (TO). To the best of our knowledge\, 
 we obtain the weakest conditions for their recovery when TO is known. Base
 d on this\, we design a two-stage causal generative model that first infer
 s the causal order from observations in a zero-shot manner\, thus by-passi
 ng the search\, and then learns the generative fixed-point SCM on the orde
 red variables. To infer TOs from observations\, we propose to amortize the
  learning of TOs on generated datasets by sequentially predicting the leav
 es of graphs seen during training. To learn fixed-point SCMs\, we design a
  transformer-based architecture that exploits a new attention mechanism en
 abling the modeling of causal structures\, and show that this parameteriza
 tion is consistent with our formalism. Finally\, we conduct an extensive e
 valuation of each method individually\, and show that when combined\, our 
 model outperforms various baselines on generated out-of-distribution probl
 ems. \n\nBio: Meyer Scetbon is  is a Researcher at Microsoft Research Camb
 ridge\, previously a Research Scientist Intern at Meta AI Paris. He earned
  his PhD from CREST - ENSAE\, Institut Polytechnique de Paris in April 202
 3\, supervised by Marco Cuturi. His research interests include foundation 
 models\, particularly their integration of causality with observable world
  understanding. His doctoral work focused on applying optimal transport to
  large-scale Machine Learning challenges and related applications.\n\n\nSp
 eaker: Wanru Zhao (University of Cambridge)\nTitle: Decentralised Collabor
 ative Large Language Model: Multilingual Applications and Data Quality Con
 trol\n5:45 pm - 6:00 pm\nAbstract: Recent research has highlighted the imp
 ortance of datasets in scaling large language models (LLMs)\; however\, da
 ta is spread across silos and locations in different formats and is hard t
 o find. This talk covers our recent works which explore decentralised coll
 aborative development of LLMs as a solution for data-sharing constraints\,
  which could benefit the broader public in the era of LLMs. Automated data
  quality control faces unique challenges in collaborative settings where d
 ata cannot be directly shared between different silos. To tackle this issu
 e\, we propose a novel data quality control technique based on training dy
 namics to enhance the quality of data from different private domains in th
 e collaborative training setting\, such as model merging. \n\nBio:\nWanru 
 Zhao is a PhD student in the Department of Computer Science and Technology
  at the University of Cambridge. Her research currently focuses on decentr
 alized collaborative development of modular large language models with che
 aply communicable updates and data attribution.
LOCATION:The Auditorium\, 21 Station Rd
END:VEVENT
END:VCALENDAR