Events at Physics

<< Fall 2020 Spring 2021 Summer 2021 >>
Subscribe your calendar or receive email announcements of events

Events on Wednesday, April 7th, 2021

Physics ∩ ML Seminar
A Mathematical Exploration of Why Language Models Help Solve Downstream Tasks
Time: 11:00 am - 12:15 pm
Place: Online Seminar: Please sign up for our mailing list at for zoom link
Speaker: Nikunj Saunshi, Princeton University
Abstract: Autoregressive language models, pretrained on large text corpora to do well on next word prediction, have been successful at solving many downstream tasks, even with zero-shot usage. However, there is little theoretical understanding of this success. We initiate a mathematical study of this phenomenon for the downstream task of text classification by considering the following questions: (1) What is the intuitive connection between the pretraining task of next word prediction and text classification? (2) How can we mathematically formalize this connection and quantify the benefit of language modeling? For (1), we hypothesize, and verify empirically, that classification tasks of interest can be reformulated as sentence completion tasks, thus making language modeling a meaningful pretraining task. With a mathematical formalization of this hypothesis, we make progress towards (2) and show that language models that are \epsilon-optimal in cross-entropy (log-perplexity) learn features that can linearly solve such classification tasks with O(\sqrt{\epsilon})-error, thus demonstrating that doing well on language modeling can be beneficial for downstream tasks. We experimentally verify various assumptions and theoretical findings, and also use insights from the analysis to design a new objective function that performs well on some classification tasks.
Host: Gary Shiu
Add this event to your calendar
Department Meeting
Department Meeting - CANCELLED
Time: 12:15 pm - 1:15 pm
Place: cancelled today due to lack of urgent business
Speaker: Sridhara Dasu, Department Chair
Host: Sridhara Dasu
Add this event to your calendar