IJS KOLOKVIJ, četrtek, 29. 6. 2023, ob 13. uri, prof. dr. Thomas G. Dietterich

[Nataša Goševac]

Vabimo vas na 14. predavanje iz sklopa »Kolokviji na IJS« v akademskem letu 2022/23, ki bo v četrtek (ne sredo!), 29. junija 2023, ob 13:00 v Veliki predavalnici Instituta »Jožef Stefan« na Jamovi cesti 39 v Ljubljani. Predavanje bo v angleščini. Na voljo bodo kava, pijača in prigrizki.

 

 

prof. dr. Thomas G. Dietterich

Oregon State University, ZDA

 

Kaj je narobe z velikimi jezikovnimi modeli in s čim bi jih morali nadomestiti

 

Veliki jezikovni modeli (LLM, iz angleščine) zagotavljajo prednaučeno osnovo za učenje številnih zanimivih sistemov umetne inteligence. Vendar imajo veliko pomanjkljivosti. Njihovo učenje in posodabljanje je drago, imajo slabo nejezikovno znanje, dajejo napačne in nasprotujoče si izjave, ki so za nameček lahko tudi družbeno in etično neprimerne. V predavanju bom povzel te pomanjkljivosti in trenutna prizadevanja za njihovo reševanje znotraj obstoječega okvira LLM-jev. Nato se bom zavzel za drugačno, bolj modularno arhitekturo, ki razgradi funkcije obstoječih LLM-jev in jim doda več drugih komponent. Verjamem, da lahko ta alternativa odpravi vse pomanjkljivosti LLM-jev. Razmišljal bom o tem, kako bi to modularno arhitekturo lahko zgradili s kombinacijo strojnega učenja in inženiringa.

 

O predavatelju: Thomas G. Dietterich je zaslužni profesor na Fakulteti za elektrotehniko in računalništvo na Državni univerzi Oregon v ZDA. Je eden od pionirjev na področju strojnega učenja ter je avtor več kot 225 recenziranih člankov in dveh knjig. Njegove trenutne raziskovalne teme vključujejo robustno umetno inteligenco, robustne sisteme človek-AI in aplikacije na področju trajnosti.

 

 

Napovednik predavanja najdete tudi na naslovu  <http://www.ijs.si/ijsw/Koledar_prireditev> http://www.ijs.si/ijsw/Koledar_prireditev, posnetke preteklih predavanj pa na  <https://video.arnes.si/?hashtag=Kolokviji_IJS> Arhiv TV IJS in  <http://videolectures.net/kolokviji_ijs> http://videolectures.net/kolokviji_ijs.

 

 

--

 

 

We invite you to the 14th lecture in the framework of »JSI Colloquia« in the academic year 2022/23, which will take place on Thursday (not Wednesday!), June 29, 2023, at 1:00 PM in the main lecture hall of Jožef Stefan Institute, Jamova cesta 39 in Ljubljana. The lecture will be held in English. Coffee, drinks and snacks will be available.

 

 

Prof. dr. Thomas G. Dietterich

Oregon State University, USA

 

What's wrong with large language models and what we should be building instead

 

Large Language Models (LLMs) provide a pre-trained foundation for training many interesting artificial intelligence systems. However, they have many shortcomings. They are expensive to train and to update, their non-linguistic knowledge is poor, they make false and self-contradictory statements, and these statements can be socially and ethically inappropriate. In this talk, I will review these shortcomings and current efforts to address them within the existing LLM framework. I will then argue for a different, more modular architecture that decomposes the functions of existing LLMs and adds several additional components. I believe this alternative can address all of the shortcomings of LLMs. I will speculate about how this modular architecture could be built through a combination of machine learning and engineering.

 

About the lecturer: Thomas G. Dietterich is Distinguished Professor Emeritus in the School of Electrical Engineering and Computer Science at Oregon State University, USA. He is one of the pioneers of the field of machine learning and has authored more than 225 refereed publications and two books. His current research topics include robust artificial intelligence, robust human-AI systems, and applications in sustainability.

 

 

You can find the lecture trailer at  <http://www.ijs.si/ijsw/Koledar_prireditev> http://www.ijs.si/ijsw/Koledar_prireditev, and recordings of past lectures at  <https://video.arnes.si/?hashtag=Kolokviji_IJS> Arhiv TV IJS and  <http://videolectures.net/kolokviji_ijs> http://videolectures.net/kolokviji_ijs.

 

 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.ijs.si/pipermail/kolokvij-ijs/attachments/20230622/b38619f3/attachment.htm>