Cinepile: A long video question answering dataset and benchmark Best Paper Award, CVPR 2024 Synth4CV
This paper introduces a dataset and benchmark for long video question answering.
Publications
For the most up-to-date list, check my Google Scholar profile.
Preprints
Conference Papers
This paper explores how natural language representations and different types of hints influence end-user debugging accuracy in AI-assisted programming.
A framework to compare NLP models by analyzing their shared invariances to linguistic perturbations, offering insights into model evolution and performance.
Pretraining induces transferable invariances, which are retained in shallow layers and compressed during finetuning, shedding light on why pretrained models excel.
DE-CROP enhances certified robustness of pretrained models with limited data by generating diverse samples and optimizing denoiser training in logit space.
A novel test-time adversarial defense method that detects and corrects adversarial samples without requiring training data, significantly improving robustness.
This paper presents a method for visible-thermal person re-identification.
Journal Articles
A simple yet effective self-distillation approach enhances adversarial robustness in few-shot learning without requiring adversarial samples, achieving significant improvements with minimal computational overhead.