Výsledky vyhledávání - "Sinha, Aradhana"

Report

Automated Adversarial Discovery for Safety Classifiers

Autor: Lal, Yash Kumar, Lahoti, Preethi, Sinha, Aradhana, Qin, Yao, Balashankar, Ananth

Safety classifiers are critical in mitigating toxicity on online forums such as social media and in chatbots. Still, they continue to be vulnerable to emergent, and often innumerable, adversarial attacks. Traditional automated adversarial data genera

Externí odkaz: http://arxiv.org/abs/2406.17104

Zobrazit plný text záznamu

Report

Generalized People Diversity: Learning a Human Perception-Aligned Diversity Representation for People Images

Autor: Srinivasan, Hansa, Schumann, Candice, Sinha, Aradhana, Madras, David, Olanubi, Gbolahan Oluwafemi, Beutel, Alex, Ricco, Susanna, Chen, Jilin

Capturing the diversity of people in images is challenging: recent literature tends to focus on diversifying one or two attributes, requiring expensive attribute labels or building classifiers. We introduce a diverse people image ranking method which

Externí odkaz: http://arxiv.org/abs/2401.14322

Zobrazit plný text záznamu

Report

Improving Few-shot Generalization of Safety Classifiers via Data Augmented Parameter-Efficient Fine-Tuning

Autor: Balashankar, Ananth, Ma, Xiao, Sinha, Aradhana, Beirami, Ahmad, Qin, Yao, Chen, Jilin, Beutel, Alex

As large language models (LLMs) are widely adopted, new safety issues and policies emerge, to which existing safety classifiers do not generalize well. If we have only observed a few examples of violations of a new safety rule, how can we build a cla

Externí odkaz: http://arxiv.org/abs/2310.16959

Zobrazit plný text záznamu

Report

Break it, Imitate it, Fix it: Robustness by Generating Human-Like Attacks

Autor: Sinha, Aradhana, Balashankar, Ananth, Beirami, Ahmad, Avrahami, Thi, Chen, Jilin, Beutel, Alex

Publikováno v: Transactions on Machine Learning Research (2024)

Real-world natural language processing systems need to be robust to human adversaries. Collecting examples of human adversaries for training is an effective but expensive solution. On the other hand, training on synthetic attacks with small perturbat

Externí odkaz: http://arxiv.org/abs/2310.16955

Zobrazit plný text záznamu

Dissertation/ Thesis

Scalable black-box model explainability through low-dimensional visualizations

Autor: Sinha, Aradhana

Thesis: M. Eng., Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science, 2017.
This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and

Externí odkaz: http://hdl.handle.net/1721.1/113109

Zobrazit plný text záznamu

Vyhledávací nástroje:

Upřesnit hledání