SUNMONTUEWEDTHUFRISAT
Conferences, Lectures, & Seminars
Events for October
-
NL Seminar-Acquiring and Understanding Cross Task Generalization with Diverse NLP Tasks
Thu, Oct 06, 2022 @ 11:00 AM - 12:00 PM
Information Sciences Institute
Conferences, Lectures, & Seminars
Speaker: Qinyuan Ye, USC/ISI
Talk Title: Acquiring and Understanding Cross Task Generalization with Diverse NLP Tasks
Series: NL Seminar
Abstract: REMINDER
Meeting hosts only admit guests that they know to the Zoom meeting. Hence, you are highly encouraged to use your USC account to sign into Zoom.
If you are an outside visitor, please inform us at nlg DASH seminar DASH host AT isi DOT edu beforehand so we will be aware of your attendance and let you in.
In person attendance will be permitted for USC ISI faculty, staff, students only. Open to the public virtually via the zoom link and online.
Humans can learn and perform a new language task more efficiently than machines, when they are provided with either task instructions or only a few examples for that task. We believe such learning efficiency is partly achieved by accumulating past learning experiences, i.e., learning to learn with previously seen tasks. We refer to such capability as cross task generalization and envision it to be an integral piece towards generalist NLP systems.
In this talk, I will present our recent efforts in acquiring and understanding cross task generalization with diverse NLP tasks 1. To build a learning environment for acquiring and evaluating cross-task generalization, we construct NLP Few shot Gym, a repository of 160 few shot tasks collected from open access NLP datasets, converted to a unified text to text format, and covering diverse formats, goals and domains. We further introduce CrossFit, a few shot learning challenge that systematically evaluates an algorithms ability to quickly learn new tasks. With these resources, we conduct an empirical analysis with multi task learning and meta learning approaches, which provides fresh insights. 2. To better understand how models learn transferable skills to achieve cross task generalization, we develop task level mixture of expert models that explicitly emulates the behavior of accumulating skills and recomposing them when encountering a new task. Our empirical results suggest that training task level mixture of experts can alleviate negative transfer and achieve better few shot performance on unseen tasks; further we find that the learned routing decisions and experts partially rediscover human categorization of NLP tasks.
Biography: Qinyuan Ye is a fourth year CS Ph.D. student at University of Southern California, advised by Prof. Xiang Ren. Her research interests lie in natural language processing. In particular she is interested in approaches that reduce human annotation efforts, including methods leveraging distant supervision, high level human supervision for example explanations, instructions, and metalearning. Prior to USC, she was an undergraduate student at Tsinghua University, majoring in Automation.
Host: Jon May and Meryem M'hamdi
More Info: https://nlg.isi.edu/nl-seminar/
Webcast: https://www.youtube.com/watch?v=hpIohClvinsLocation: Information Science Institute (ISI) - Virtual Only
WebCast Link: https://www.youtube.com/watch?v=hpIohClvins
Audiences: Everyone Is Invited
Contact: Pete Zamar
Event Link: https://nlg.isi.edu/nl-seminar/
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor. -
NL Seminar- Understanding and Improving Learning through Inference with Large Language Models
Thu, Oct 20, 2022 @ 11:00 AM - 12:00 PM
Information Sciences Institute
Conferences, Lectures, & Seminars
Speaker: Sewon Min, University of Washington
Talk Title: Understanding and Improving Learning through Inference with Large Language Models
Series: NL Seminar
Abstract: THIS TALK WILL NOT BE RECORDED, IT WILL BE BROADCAST LIVE ONLY*
REMINDER:
Meeting hosts only admit guests that they know to the Zoom meeting. Hence, you are highly encouraged to use your USC account to sign into Zoom.
If you are an outside visitor, please inform us at nlg DASH seminar DASH host AT isi DOT edu beforehand so we will be aware of your attendance and let you in.
In person attendance will be permitted for USC ISI faculty, staff, students only. Open to the public virtually via the zoom link and online.
Language models are capable of learning at inference also referred to as in context learning, learning a new task by conditioning on k examples and making a prediction for a new input with no parameter updates. While impressive, models suffer from high variance and low worst case accuracy. Moreover, we do not understand how or why in context learning works. In the first part of the talk, I will introduce new methods that lead to significant performance gains by reducing variance and improving worst case accuracy. I will present a new inference method as well as a new training method, of which combination enables the model to outperform a 230x bigger language model. In the second part of the talk, I will show that in context learning in fact works very differently from conventional learning: the model does not benefit from the correctly paired training data, but rather benefit from the correct specification of the independent distribution of inputs and labels. Finally, I will conclude the talk with lessons learned, limitations and avenues for future work.
Biography: Sewon Min is a Ph.D. student in the Paul G. Allen School of Computer Science and Engineering at the University of Washington, advised by Prof. Luke Zettlemoyer and Prof. Hannaneh Hajishirzi. She is also a part time visiting researcher at Meta AI. Her research is in the area of natural language processing and machine learning. Her work specifically focuses on question answering, natural language understanding, knowledge representation and building general purpose language understanding models. She is a recipient of the 2022 JP Morgan Ph.D. Fellowship. She has co organized multiple workshops and tutorials at ACL, EMNLP, NeurIPS and AKBC, including a workshop on Machine Reading for Question Answering, a competition on Efficient Open domain Question Answering, a workshop on Representation Learning for NLP, workshop on Semiparametric Methods in NLP, and a tutorial on Zero and Few shot Learning with Pretrained Language Models. Prior to UW, she obtained a B.S. degree in Computer Science & Engineering from Seoul National University.
Host: Jon May and Meryem M'hamdi
More Info: https://nlg.isi.edu/nl-seminar/
Webcast: https://usc.zoom.us/j/94452797669Location: Information Science Institute (ISI) - Virtual and ISI-Conf Rm#689
WebCast Link: https://usc.zoom.us/j/94452797669
Audiences: Everyone Is Invited
Contact: Pete Zamar
Event Link: https://nlg.isi.edu/nl-seminar/
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor. -
NL Seminar
Thu, Oct 27, 2022 @ 11:00 AM - 12:00 PM
Information Sciences Institute
Conferences, Lectures, & Seminars
Speaker: Eric Wallace, University of Cal-Berkeley
Talk Title: Emerging Vulnerabilities in Large-scale NLP Models
Series: NL Seminar
Abstract: Abstract: REMINDER
Meeting hosts only admit guests that they know to the Zoom meeting. Hence, you are highly encouraged to use your USC account to sign into Zoom.
If you are an outside visitor, please inform us at nlg DASH seminar DASH host AT isi DOT edu beforehand so we will be aware of your attendance and let you in.
In person attendance will be permitted for USC ISI faculty, staff, students only. Open to the public virtually via the zoom link and online.
The current era of machine learning and natural language processing is dominated by scale modern models use supermassive parameter counts, dataset sizes, and compute budgets. While this scaling undoubtedly unlocks new capabilities and performance improvements, it may also expose new vulnerabilities, risks, and harms. In this talk, I will discuss a series of vulnerabilities that emerge in large scale NLP models that not only expose worrisome security and privacy risks, but also provide new perspectives into how and why the models work. Concretely, I will show how adversaries can extract private training data, steal model weights, and poison training sets, all using limited black box access to models. Throughout the talk, I'll provide a particular focus on insights that we can derive from these attacks, especially regarding the impact of model scaling.
Biography: Eric Wallace is a 4th year PhD student at UC Berkeley advised by Dan Klein and Dawn Song. His research interests focus on making large language models more robust, trustworthy, secure, and private. Eric's work is supported by the Apple Fellowship in AI/ML, and he received the best demo award at EMNLP 2019.
Host: Jon May and Meryem M'hamdi
More Info: https://nlg.isi.edu/nl-seminar/
Webcast: https://www.youtube.com/watch?v=42LNH1dTlggLocation: Information Science Institute (ISI) - Virtual and ISI-Conf Rm#689
WebCast Link: https://www.youtube.com/watch?v=42LNH1dTlgg
Audiences: Everyone Is Invited
Contact: Pete Zamar
Event Link: https://nlg.isi.edu/nl-seminar/
This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor.
