BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//wp-events-plugin.com//7.4.3//EN
TZID:Asia/Kolkata
X-WR-TIMEZONE:Asia/Kolkata
BEGIN:VEVENT
UID:228@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20260930T113000
DTEND;TZID=Asia/Kolkata:20260930T123000
DTSTAMP:20260916T172535Z
URL:https://cds.iisc.ac.in/events/ph-d-thesis-colloquium-102-cds-30-septem
 ber-2026-understanding-measuring-and-mitigating-bias-in-visual-recognition
 -and-generation/
SUMMARY:Ph.D: Thesis Colloquium: 102: CDS: 30\, September 2026 "Understandi
 ng\, Measuring\, and Mitigating Bias in Visual Recognition and Generation"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nPh.D. Thesis Col
 loquium\n\n\n\nSpeaker: Ms. Abhipsa Basu\nS.R. Number: 06-18-01-10-12-20-2
 -19109\nTitle: "Understanding\, Measuring\, and Mitigating Bias in Visual 
 Recognition and Generation"\nResearch Supervisor: Prof. Venkatesh Babu\nDa
 te &amp\; Time : September 30\, 2026 (Wednesday)\,11:30 AM\nVenue : #102\,
  CDS Seminar Hall\n\n\n\nABSTRACT\nHuman beings form their understanding o
 f the world through the people\, places\, and events they encounter\, but 
 these experiences are inherently limited and uneven\, causing our percepti
 ons to reflect harmful biases and stereotypes shaped by our surroundings. 
 Artificial intelligence systems\, trained on data collected from the world
 \, can inherit and amplify these same biases. In discriminative systems\, 
 this can lead models to rely on spurious associations or perform poorly fo
 r underrepresented groups\; in generative systems\, it can result in stere
 otypical or narrow representations of people\, places\, and environments. 
 Because the datasets used to train modern AI systems are collected at scal
 e and are difficult to curate or control\, dedicated methods are needed to
  understand\, measure\, and mitigate such biases. This thesis studies thes
 e challenges in visual AI.\n\nThe first part investigates and mitigates di
 fferent forms of bias in discriminative vision systems. For image classifi
 cation\, we address biases arising from imbalanced data distributions and 
 spurious correlations through an adaptive clustering-based margin loss tha
 t enables debiasing in the presence of frozen or blackbox feature extracto
 rs\, and explore diffusion-based synthetic data generation for constructin
 g more balanced training sets. Extending beyond unimodal image classificat
 ion\, we investigate language biases in visual question answering (VQA)\, 
 where spurious question-answer correlations can cause models to ignore ima
 ge content. We mitigate these biases through a learnable adaptive margin-l
 oss framework that preserves both in-distribution and out-of-distribution 
 performance.\n\nThe second part examines bias in text-to-image (T2I) gener
 ation from a geographic perspective. Through a large-scale crowdsourced st
 udy\, we find that T2I models tend to represent everyday visual concepts u
 sing imagery associated with only a limited set of countries when geograph
 ic information is unspecified in the prompt\, while many other regions rem
 ain poorly represented. Explicitly specifying a country substantially impr
 oves its representation\, suggesting that models can depict diverse region
 s but do not necessarily choose to do so by default. We further investigat
 e the training data underlying these models by geographically profiling ca
 ptions from large-scale vision-language datasets and find that a large maj
 ority of image-caption pairs can be geographically attributed to only a fe
 w countries.\n\nGeographical representation\, however\, is not only a ques
 tion of how frequently a country appears\, but also of how it is portrayed
 . We introduce GeoDiv\, an interpretable framework for measuring diversity
  along visual and socioeconomic dimensions\, revealing that countries are 
 often represented through a limited range of characteristics. Finally\, we
  investigate the mechanistic origins of these socioeconomic biases in the 
 T2I pipeline and find that they are likely distributed across multiple com
 ponents rather than arising from a single source.\n\nTogether\, these work
 s advance the diagnosis and mitigation of bias in modern vision systems\, 
 providing methods and frameworks for building more equitable visual AI.\n\
 n\n\n ALL ARE WELCOME
CATEGORIES:Events,Ph.D. Thesis Colloquium
END:VEVENT
BEGIN:VTIMEZONE
TZID:Asia/Kolkata
X-LIC-LOCATION:Asia/Kolkata
BEGIN:STANDARD
DTSTART:20250930T113000
TZOFFSETFROM:+0530
TZOFFSETTO:+0530
TZNAME:IST
END:STANDARD
END:VTIMEZONE
END:VCALENDAR