Mon, Jun 13, 2022 @ 02:00 PM - 03:00 PM
Information Sciences Institute
Conferences, Lectures, & Seminars
Speaker: Naomi Saphra, NYU
Talk Title: Sources of Variance in Pretraining and Finetuning LLMs
Series: NL Seminar
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 DOR edu beforehand so we will be aware of your attendance and let you in.
You have engaged in the very modern practice of transfer learning. You pretrained a model on a self supervised objective, then you finetuned it on a downstream task, and you find excellent performance on the test set. Aha, you say. I found a good pretraining procedure. Did you? You try finetuning again. The results are terrible! Aha, you say. I found a bad finetuning procedure. Did you?
The random seeds for both pretraining and finetuning stages have a substantial influence on outcome. However, it is computationally expensive to pretrain new models, so measuring the robustness of a procedure across different seeds can be prohibitive. This talk will address, first, the influence that a pretraining seed has on both in domain and OOD performance. Then we will address the role of the finetuning seed. Much variation in OOD generalization can be ascribed to where the finetuning seeds direct SGD trajectories. In particular, we discuss how to predict generalization behavior in a finetuned model, based on topographic properties of its region of the loss surface. By understanding the degree of influence that random seeds have on performance, we can fairly evaluate a robust training procedure, rather than a single set of parameters. By understanding the mechanism of that influence, we can go further by developing improved training methods.
Biography: Naomi has interests relating to NLP learning dynamics how models learn to encode linguistic structure, and how we can encode useful inductive biases into the training process. Having earned a PhD from University of Edinburgh, they are now a postdoc at NYU. In their spare time, they play roller derby under the name Gaussian Retribution, do standup comedy, and shepherd programmers who cannot type into the world of code dictation.
Host: Jon May and Thamme Gowda
More Info: https://nlg.isi.edu/nl-seminar/
WebCast Link: https://www.youtube.com/watch?v=Lni4PIlbJjI
Audiences: Everyone Is Invited
Contact: Pete Zamar
Event Link: https://nlg.isi.edu/nl-seminar/