“Less is More — Toward Efficient Vision Transformers in Perception and Reasoning” Seminar

Join us for the CS & CIVS Distinguished Speaker Seminar presented by Professor Yang Ni.

The remarkable success of recent large foundation models stems from the versatility of the transformer architecture and large-scale pre-training on massive multimodal datasets. Originally introduced in natural language processing, the attention-centric transformer design has become the dominant paradigm across domains.

More recently, transformers have emerged as the backbone of modern computer vision models. However, when high-resolution images are processed, they are decomposed into thousands of tokens, orders of magnitude more than typical text inputs, causing the quadratic computational complexity.

To address this challenge, current research has focused on reducing the number of image tokens participating in attention. A central question then arises: which tokens truly carry essential information, and which can be safely pruned?

This talk provides an overview of recent advances in token reduction strategies for efficient transformers, with an emphasis on large vision-language models (LVLMs) where perception and reasoning are deeply integrated.

Location

CIVS Theater, Powers Building
Hammond Campus

To request a disability-related accommodation, please contact the Office for Civil Rights at civilrights@pnw.edu or (219) 989-2163 five days prior to the event.

In accordance with Purdue policies, all persons have equal access to Purdue University’s educational programs, services, and activities, without regard to race, religion, color, sex, age, national origin or ancestry, genetic information, marital status, parental status, sexual orientation, gender identity and expression, disability, or status as a veteran. See Purdue’s Nondiscrimination Policy Statement. If you have any questions or concerns regarding these policies, please contact the Office for Civil Rights at civilrights@pnw.edu or (219) 989-2337.