Caption: 

A research poster titled "Can we find a universal pruned VLM?" is presented, highlighting issues and solutions in Vision-Language Model (VLM) pruning. Co-authored by Matteo Farina and Massimiliano Patacchiola from Università di Trento and sponsored by Cisco, Università di Trento, and Fondazione Bruno Kessler, the poster discusses problems such as speed and practicality in current VLM pruning practices. Diagrams illustrate the difference between Task-Specific VLM Pruning and Task-Agnostic Vision-Language Pruning, showcasing how dense VLMs can be pruned for various applications, from captioning to retrieval. The research aims to enhance efficiency and adaptability in managing VLMs across diverse tasks. 

Credits to the presented institutions and their logos: Cisco, Università di Trento, and Fondazione Bruno Kessler.
Text transcribed from the image:
CISCO
UNIVERS
RSITAS
ATHE
UNIVERSITÀ
DI TRENTO
FONDAZIONE
BRUNO KESSLER
1 Some problems in VLM Pruning
Speed: Prior works focus on gradually pruning during
training. Practicality: one must re-prune whenever the
downstream task changes.
Can we find a universal
pruned VLM?
Task-Specific VLM Pruning (current)
t₁
CAPTIONING
Dense VLM
YAMAHA
"a brown tower
with a clock on top."
"A plate of food and
a glass of liquid."
"a cat lying down
on a bicycle seat."
Pruned
VLM
t2
RETRIEVAL
Pruned
VLM
VQA
"How many street
lights do you see?"
"One."
Pruned
VLM
Task-Agnostic Vision-Language Pruning (ours)
Dense VLM
"a group of buildings
under nice blue sky."
Pruned
VLM
CA
MULTIFLOW: Shiftin
Matteo Farina, Massimilia
(1) University
D
"Giraffes graze on
ow shrubs under
9
Dense
VLM