Nihal Murali

I’m a fifth-year Ph.D. candidate in the Intelligent Systems Program at the University of Pittsburgh, jointly advised by Kayhan Batmanghelich (BatmanLab, Boston University) and Shyam Visweswaran (VisLab). My research lies at the intersection of machine learning, medicine, and human–AI collaboration—developing algorithms that make AI systems safer, more interpretable, and clinically reliable.

Broadly, I study how AI models behave under uncertainty and how they interact with human expertise in high-stakes decision making. My work spans from understanding model failures such as shortcut and spurious learning to designing hybrid human–AI systems that integrate algorithmic predictions with human judgment for expert-level performance in healthcare applications.

“When should AI decide, and when should it defer?
How can human expertise be modeled and trusted in the loop?”

My broader goal is to advance AI systems that understand their limitations and collaborate effectively with humans—especially in domains where reliability and safety are non-negotiable.

I was a visiting researcher jointly collaborating with the Machine Learning Research Group (MLRG) at the University of Guelph (Ontario, Canada) and the University of Toronto, advised by Prof. Graham W. Taylor and Prof. Joel D. Levine, on computer-vision models for biological data.

Previously, I interned at the Robotics Research Center (RRC) at the International Institute of Information Technology, Hyderabad, where I worked with Prof. Madhav Krishna and Dr. Krishna Murthy on robotics and perception.

I received my M.E. in Software Systems and B.E. (Hons.) in Electrical and Electronics Engineering from BITS Pilani, where I was advised by Prof. Surekha Bhanot.

“I build AI systems that know when to ask for help.”

Email / CV / Scholar / GitHub

	Hybrid novice–AI system achieves expert-level performance in intraoperative ischemia detection Nihal Murali, Amir I. Mina, Harsh Sinha, Joshua W. Anderson, Yash Raka, Hung-Ching Chang, Homa K. Amiri, Parthasarathy D. Thirumala, Kayhan Batmanghelich^†, Shyam Visweswaran^† Under review, npj Digital Medicine (2026) medRxiv preprint ^†These authors contributed equally as corresponding authors Paper / Code A hybrid novice–AI system for detecting cerebral ischemia from continuous intraoperative EEG during carotid endarterectomy. By dynamically weighting a minimally-trained novice monitor's assessment against an AI model's prediction, the system achieves performance statistically non-inferior to board-certified neurophysiologists—halving the AI model's false-positive rate at 80% sensitivity while improving precision and calibration.
	Machine learning to detect intraoperative ischemia from electroencephalography in carotid endarterectomy surgery Shyam Visweswaran^, Mehdi Nourelahi^, Amir I. Mina, Jessi U. Espino, Nihal Murali, Kayhan Batmanghelich, Parthasarathy D. Thirumala Under review, npj Health Systems (2026) ^*Equal contribution Paper A comparative study of supervised machine learning models—random forests, gradient boosting, logistic regression, SVMs, and naive Bayes—for automated detection of cerebral ischemia from quantitative EEG during carotid endarterectomy. Random forests achieved the highest sensitivity and precision–recall performance, with alpha-band activity and interhemispheric asymmetry emerging as the most discriminative predictors—pointing to ML as a scalable aid to expert neuromonitoring.
	Beyond Distribution Shift: Spurious Features Through the Lens of Training Dynamics Nihal Murali, Aahlad Puli, Ke Yu, Rajesh Ranganath, Kayhan Batmanghelich TMLR 2023, ICMLW 2023 Paper / Github / Video / Poster / Slides / Talk Reframes spurious-feature learning through the lens of a network's training dynamics rather than distribution shift alone. The key insight: not all spurious features are harmful—a shortcut only hurts generalization when it is easier to learn than the core signal, a distinction that is model- and dataset-dependent. Using example-difficulty metrics such as Prediction Depth, the paper shows that these harmful shortcuts surface in the learning dynamics of a network's early layers, making them detectable early in training—before they degrade deployment performance—and enabling timely intervention. Validated on both simulated and real medical and vision datasets.
	Augmentation by Counterfactual Explanation – Fixing an Overconfident Classifier Sumedha Singla^, Nihal Murali^, Forough Arabshahi, Sofia Triantafyllou, Kayhan Batmanghelich WACV 2023 ^*Equal contribution Paper / GitHub Proposed a counterfactual-based fine-tuning method (ACE) that uses GAN-generated augmentations to reduce overconfidence and improve uncertainty estimation in deep neural networks.
	Classification and re-identification of fruit fly individuals across days with convolutional neural networks Nihal Murali, Jon Schneider, Joel Levine, Graham Taylor WACV 2019 Paper / GitHub / Dataset Developed a CNN-based method to maintain individual fruit fly identities across days, enabling long-term behavioral tracking without physical tagging.
	Can Drosophila melanogaster tell who’s who? Jon Schneider, Nihal Murali, Graham Taylor, Joel Levine PLOS ONE 2018 Paper / News Demonstrated that individual fruit flies possess unique visual features detectable by deep convolutional networks—suggesting visual discrimination among conspecifics despite low-resolution vision.
	Classification and Re-Identification of Fruit Flies using Deep Convolutional Networks Bachelor’s Thesis, BITS Pilani Developed a deep learning system using ResNet architectures to automatically recognize and re-identify individual Drosophila melanogaster across days—eliminating manual tagging and enabling long-term behavioral experiments.

Template taken from Jon Barron