Portfolio item number 1
Short description of portfolio item number 1
Posts by Collection
portfolio
Portfolio item number 2
Short description of portfolio item number 2
publications
Generative Adversarial Masking (Student Paper)
Published in International Student Conference on Artificial Intelligence, 2021
A novel technique for efficient regularization in supervised learning using GANs for adversarial masking using semantic information.
Recommended citation: Prashant, Mohit. "Generative Adversarial Masking" International Student Conference on Artificial Intelligence. 2021.
Download Paper
PAC-Based Formal Verification for Out-of-Distribution Data Detection
Published in IEEE International Conference on System Reliability and Safety, 2022
We present a technique for detecting OOD instances in autonomous driving with PAC guarantees.
Recommended citation: Prashant, Mohit, and Arvind Easwaran. "PAC-Based Formal Verification for Out-of-Distribution Data Detection." 2022 6th International Conference on System Reliability and Safety (ICSRS). IEEE, 2022.
Download Paper
Guaranteeing Out-Of-Distribution Detection in Deep RL via Transition Estimation
Published in Proceedings of the AAAI Conference on Artificial Intelligence, 2025
We provide a basis for defining out-of-distribution execution in RL and provide a framework for detection with guarantees.
Recommended citation: Prashant, Mohit, et al. "Guaranteeing Out-Of-Distribution Detection in Deep RL via Transition Estimation". Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 39. No. 12. 2025.
Download Paper
Probably Approximately Correct Guarantees for Reinforcement Learning in Continuous State-Spaces Through Discretization
Published in Journal of AI Research, 2025
We provide error-confidence guarantees on policies for RL algorithms applied to discretized environments.
Recommended citation: Preprint
Download Paper
Improving Reinforcement Learning Sample-Efficiency using Local Approximation
Published in Proceedings of the ECAI Conference on Artificial Intelligence, 2025
We provide tighter sample-complexity bounds for reinforcement learning on metric spaces through local approximation and provide a respective algorithm for more efficient learning.
Recommended citation: Preprint
Download Paper
talks
Building Robust Systems Using Adversarial Input Masking
Published:
Presentation at STCAI2021 on using generative adversarial networks to build robust systems using a form of input masking. The results of the study showed that informed masking allowed for better regularization on image prediction than dropout and can be used in conjunction with existing data augmentation methods. Recording attached inside.
Efficiently Scaling Model-Free Reinforcement Learning with Local Approximation
Published:
Presented a guest lecture at UPenn, SEAS on efficient scaling of RL systems such that the sample-complexity of learning is a linear function of state-space size using localized MDP approximations, improving on SOTA log-linear scaling bounds. The full paper is available here.
Addressing Guaranteed Out-Of-Distribution Policy Execution
Published:
Presenting a brief talk on our AAAI25 paper regarding OOD policy execution and detection. Recording attached inside.
teaching
Operating Systems (Old)
Undergraduate Course, College of Computer and Data Science, NTU, 2023
Worked as a teaching assistant in preparing, teaching and assessing the undergraduate course on operating systems.
Data Structures and Algorithms (SC1007)
Undergraduate Course, College of Computer and Data Science, NTU, 2023
Worked as a teaching assistant in preparing, teaching and assessing the undergraduate course on data structures and algorithms.
Operating Systems (SC2005)
Undergraduate Course, College of Computer and Data Science, NTU, 2025
Worked as a teaching assistant in preparing, teaching and assessing the undergraduate course on operating systems. Designed the laboratory curriculum for the latest iteration of the course.
ML Safety and Learning Theory
Postgraduate Research Workshops, -, 2025
Conducting a series of online and in-person workshops on ML safety and learning theory using probably approximately correct guarantees. The aim of these workshops is to introduce the listener to the notion of error-confidence bounds within ML and demonstrate practical applications of learning theory in deriving safety guarantees for learning-enabled systems.