Jove
Visualize
Contact Us
JoVE
x logofacebook logolinkedin logoyoutube logo
ABOUT JoVE
OverviewLeadershipBlogJoVE Help Center
AUTHORS
Publishing ProcessEditorial BoardScope & PoliciesPeer ReviewFAQSubmit
LIBRARIANS
TestimonialsSubscriptionsAccessResourcesLibrary Advisory BoardFAQ
RESEARCH
JoVE JournalMethods CollectionsJoVE Encyclopedia of ExperimentsArchive
EDUCATION
JoVE CoreJoVE BusinessJoVE Science EducationJoVE Lab ManualFaculty Resource CenterFaculty Site
Terms & Conditions of Use
Privacy Policy
Policies

Related Experiment Video

Updated: Sep 18, 2025

Visualizing Visual Adaptation
04:43

Visualizing Visual Adaptation

Published on: April 24, 2017

9.1K

TDA-L: Reducing Latency and Memory Consumption of Test-Time Adaptation for Real-Time Intelligent Sensing.

Rahim Hossain1, Md Tawheedul Islam Bhuian1, Kyoung-Don Kang1

  • 1School of Computing, State University of New York at Binghamton, 4400 Vestal Parkway East, Vestal, NY 13850, USA.

Sensors (Basel, Switzerland)
|June 27, 2025
PubMed
Summary

Related Concept Videos

You might also read

Related Articles

Articles linked to this work by shared authors, journal, and citation graph.

Sort by
Same author

Corun: Concurrent Inference and Continuous Training at the Edge for Cost-Efficient AI-Based Mobile Image Sensing.

Sensors (Basel, Switzerland)·2024
Same author

Filtering Empty Video Frames for Efficient Real-Time Object Detection.

Sensors (Basel, Switzerland)·2024
See all related articles

This study introduces TDA-L, a novel framework for vision-language models that enhances real-time sensing by improving adaptability to changing conditions without backpropagation. TDA-L offers efficient, robust performance for visually impaired users by reducing computational costs.

Area of Science:

  • Computer Vision
  • Artificial Intelligence
  • Natural Language Processing

Background:

  • Vision-language models (VLMs) are crucial for real-time intelligent sensing, enabling applications like image analysis for the visually impaired.
  • Adapting VLMs to real-world distribution shifts (e.g., lighting, weather) at test time is challenging.
  • Existing test-time adaptation methods often require computationally expensive gradient-based fine-tuning, hindering real-time application.

Purpose of the Study:

  • To develop a computationally efficient and effective test-time adaptation method for VLMs.
  • To improve the robustness of VLMs against distribution shifts without compromising real-time performance.
  • To introduce TDA-L, a framework integrating Low-Rank Adaptation (LoRA) with Training-Free Dynamic Adapter (TDA).

Main Methods:

Keywords:
edge analyticslow-rank adaptationreal-time intelligent sensingtest-time adaptationvision–language models

More Related Videos

Real-Time Proxy-Control of Re-Parameterized Peripheral Signals using a Close-Loop Interface
11:54

Real-Time Proxy-Control of Re-Parameterized Peripheral Signals using a Close-Loop Interface

Published on: May 8, 2021

4.7K
Automated, Long-term Behavioral Assay for Cognitive Functions in Multiple Genetic Models of Alzheimer's Disease, Using IntelliCage
06:46

Automated, Long-term Behavioral Assay for Cognitive Functions in Multiple Genetic Models of Alzheimer's Disease, Using IntelliCage

Published on: August 4, 2018

12.3K

Related Experiment Videos

Last Updated: Sep 18, 2025

Visualizing Visual Adaptation
04:43

Visualizing Visual Adaptation

Published on: April 24, 2017

9.1K
Real-Time Proxy-Control of Re-Parameterized Peripheral Signals using a Close-Loop Interface
11:54

Real-Time Proxy-Control of Re-Parameterized Peripheral Signals using a Close-Loop Interface

Published on: May 8, 2021

4.7K
Automated, Long-term Behavioral Assay for Cognitive Functions in Multiple Genetic Models of Alzheimer's Disease, Using IntelliCage
06:46

Automated, Long-term Behavioral Assay for Cognitive Functions in Multiple Genetic Models of Alzheimer's Disease, Using IntelliCage

Published on: August 4, 2018

12.3K
  • Proposed TDA-L framework, combining TDA's dynamic key-value cache and pseudo-labeling with LoRA.
  • Applied LoRA transformations to query and cached features during inference to reduce computational overhead.
  • Evaluated TDA-L on seven diverse benchmarks to assess its performance against distribution shifts.

Main Results:

  • TDA-L maintained model accuracy compared to existing methods.
  • Achieved significantly lower latency and reduced memory consumption.
  • Demonstrated higher throughput, indicating superior efficiency for real-time applications.

Conclusions:

  • TDA-L offers a training-free, computationally efficient solution for adapting VLMs to test-time distribution shifts.
  • The integration of LoRA effectively reduces feature representation size and computational load.
  • TDA-L is well-suited for resource-constrained, real-time AI-based sensing applications.