Event Details

Speaker: Soroush Vosoughi, Dartmouth College

Abstract: In recent years, the landscape of natural language processing (NLP) has been substantially reshaped by the advent of large-scale language models (LLMs) such as BERT and GPT-3. These pre-trained models are known for achieving near-human proficiency across a variety of tasks, often requiring little to no additional training data. This performance surge is attributed partly to their extensive architectures, incorporating hundreds to thousands of millions of parameters and the vast datasets garnered from the web for pre-training. Nonetheless, these very characteristics render the models susceptible to mirroring the biases and anti-social tendencies pervasive on the web, thereby forming a significant hindrance to their application in real-world settings, particularly in social spheres. Addressing this, our lab specializes in developing methods for post hoc mitigation of such tendencies, implemented at the inference stage. This approach not only retains the prosocial inclinations during inference but also bypasses the need for expensive and untenable retraining of these models. During this talk, I will delve into a critical review of our latest endeavors aimed at refining language models to diminish bias and better align them with human ethical standards, leveraging inference-time mitigation techniques. This initiative is an essential step forward in harmonizing the functioning of language models with socially responsible and ethical objectives, thereby enhancing their reliability and applicability in addressing contemporary challenges.

REMINDER:

Meeting hosts only admit guests that they know to the Zoom meeting. Hence, you’re highly encouraged to use your USC account to sign into Zoom.

If you’re an outside visitor, please inform us at (nlg-seminar-host(at)isi.edu) beforehand so we’ll 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.

Host: POC:

Speaker Bio

Biography: Prof. Soroush Vosoughi leads the Minds, Machine, and Society group at Dartmouth. The group explores the nuances of large language models (LLMs), focusing particularly on mitigating their anti-social tendencies to foster a more responsible and transparent AI technology. His research also delves into the sphere of computational social science, creating tools that offer nuanced perspectives on various social systems and issues. Recently, their research has ventured into integrating visual data with language models, aspiring to craft a more comprehensive representation of the extensive data available, thereby inching closer to a nuanced understanding of human cognition. Prof. Vosoughi is a recipient of the Google Research Scholar Award in 2022 and an Amazon Research Award in 2019, and his work has earned several Best Paper awards and nominations, including the Outstanding Paper Award at AAAI 2021. Before joining Dartmouth, Prof. Vosoughi was a postdoctoral associate at MIT and a fellow and later an affiliate at the Berkman Klein Center at Harvard University. He received his Ph.D., MSc, and BSc from MIT in 2015, 2010, and 2008. Prof. Vosoughi's research has been supported by a variety of entities, such as the NSF, NIH, and the John Templeton Foundation.