As the computer vision research community makes rapid progress in producing algorithms with human-level performance, it is extremely critical that we take a step back and assess, and consequently promise to the consumer world, what this objective performance reported in academic literature means in the context of real-world systems and applications. As a concrete example, it is one thing for a social media organization to use an algorithm to automatically identify a person of interest in pictures uploaded to its platform. On the other hand, the use of algorithms in making life-changing decisions in areas such as healthcare (e.g., should a certain treatment be administered?) is a totally different ballgame. At the very least, the following questions will be asked of the algorithm/system by the user:

-Why is the algorithm predicting X?
-How sure is the algorithm of this prediction/decision?
-Why should I trust the algorithm?
-How can I be sure the algorithm has been fair in the process leading up to its prediction/decision?
-Is the algorithm biased?

Answers to these questions can have profound consequences depending on the application (e.g., accidents and autonomous vehicles, life/death for a patient, incarceration/freedom for an accused). Consequently, as artificial intelligence (AI) is seeing increasing adoption in a variety of daily-life applications, addressing the underlying themes of the questions above has become a matter of urgent importance. In light of these issues, we seek to provide a focused venue for academic and industry researchers and practitioners to discuss research challenges and solutions associated with learning computer vision models with the overarching requirements of fairness, data efficiency, and trustworthiness. In particular, we ask:

-How can we make our algorithms more explainable and trustworthy than they currently are?

-How can we make our algorithms more fair and less biased than they currently are?
-How can we train robust models under biased and scarce data?
-How can we detect bias or scarcity in data for a given objective function?

Topics for TCV 2022 include, but are not limited to:

-Algorithms and theories for explainable and interpretable computer vision models
-Application-specific designs for explainable computer vision, e.g., healthcare, autonomous driving, etc
-Algorithms and theories for learning computer vision models under bias and scarcity
-Performance characterization of vision algorithms and systems under bias and scarcity.
-Algorithms for secure and privacy-aware machine learning for computer vision
-Algorithms and theories for trustworthy computer vision models
-The role of adjacent fields of study (e.g, computational social science) in mitigating issues of bias and trust in computer vision

Important Dates

-Paper submission deadline: March 25, 2022 11.59pm Pacific Time
-Notification to authors: April 8, 2022 11.59pm Pacific Time
-Camera ready deadline: April 15, 2022 11.59pm Pacific Time

Workshop website 

https://fadetrcv.github.io/2022/ 

Submission Website 

https://cmt3.research.microsoft.com/TCV2022/  

Contact 

srikrishna@ieee.org, nratha@buffalo.edu

COURSE TITLE: Domain Adaptation & Generalization

LECTURER:       Vittorio Murino, vittorio.murino@univr.it; Pietro Morerio, pietro.morerio@iit.it

ORGANIZER:     University of Verona and Istituto Italiano di Tecnologia

CONTENT AND ORGANIZATION:            A standard assumption of learning based models is that training and test data share the same input distribution. However, models trained on given datasets perform poorly when tested on data acquired in different settings. This problem is known as domain shift and is particularly relevant, e.g., for visual models of agents acting in the real world or when we have no labeled data available for our target scenario. In the latter case, for instance, we could use synthetically generated data to obtain data for our target task, but this would create a mismatch between training (synthetic) and test (real) images. Filling the gap between these two different input distributions is the goal of domain adaptation (DA) algorithms. In particular, the goal of DA is to produce a model for a target domain (for which we have few or no labeled data) by exploiting labeled data available in a different, source, domain. Various DA techniques have been developed to address the domain shift problem. In this short course, we will provide an introduction to these algorithms and to domain adaptation and generalization. In particular, we will first introduce the domain shift problem, showing application scenarios where it is strongly present. Second, we will provide an overview of the algorithms that have been developed to tackle this issue. In particular, we will focus on the last research trends addressing the DA problem within deep neural networks. Lastly, we will address the domain generalization problem, which is a more challenging task because it assumes that target data is also not available, implying that the training algorithm should be devised to generalize as much as possible without any adaptation to the target in order to properly classify never observed, out-of-distribution samples.

 

REGISTRATION: Free of charge

 

WHEN: April 6, 2022 – 14.00-18.00 CET

 

WHERE: Online (link to be provided by the Lecturer after registration/enrollment)

 

HOW TO REGISTER and ENROLL: 

Both AIDA and non-AIDA students are encouraged to participate in this short course. 

 

If you are an AIDA Student* already, please: 

Step (a): Register in the course by sending an email to pietro.morerio[at]iit.it for your registration. 

AND 

Step (b): Enroll in the same course in the AIDA system using the “Enroll on this Course” button, which you can find here, so that this course enters your AIDA Certificate of Course Attendance. 

 

If you are not an AIDA Student do only step (a). 

 

*The International AI Doctoral Academy (AIDA) has 73 members, which are top AI Universities, Research centers and Industries: https://www.i-aida.org/