Usman Roshan
Ph.D. Computer Science, The University of Texas at Austin, Advisor: Tandy Warnow
Phone: 973-596-2872 (office) Email: usman@njit.edu Research
Medical Agentic AI systems
Evaluation of Medical Vision Language Models HuluMed and MedGemma, and general purpose chatbots Gemma 3, ChatGPT Plus, and Claude Pro on real previously unseen wound images
(submitted, ArXiv)
Medi-Gemma: A Hybrid Clinical Decision Support System Integrating Deterministic EMR Analytics and Retrieval-Augmented Generation
(submitted, ArXiv)
Adversarial attack challenge on CIFAR10 pairwise binary classification - we invite you to perform
white box attacks on our sign activation 01 loss neural network models.
In order to succeed in this challenge you must produce adversaries
within L-infinity distance of 16/255 of the original image using any method of your choice. Your model must
have at least 80% mean clean accuracy on pairwise CIFAR10 classification (shown in Table 1 below).
Your adversaries must bring down the model accuracy to almost 0%, which is on-par accuracy with the same
model structure with relu activations and cross-entropy loss (shown in Figure 5.5 below).
