Hi, I'm Pranav.

I'm a soon-to-graduate PhD student at the University of Oxford, advised by Prof. Ben Seymour (NDCN, IBME) and co-advised by Prof. Ioannis Havoutis (ORI). And I'm actively looking for postdoctoral research opportunities!

My background is in engineering and I'm a Neuro-AI researcher working on fundamental problems with applications to safe AI and AI for healthcare, with a keen interest in continual learning and understanding LLMs using the lens of computational neuroscience. I previously interned at Nvidia where I worked on scalable distributed deep learning.

I've been fortunate to collaborate with and learn from Peter Dayan, Rafal Bogacz, Boris Gutkin, Flavia Mancini, Sang Wan Lee, and my advisors, Ben and Ioannis.

I'm introverted. I'm a cheerful pessimist. I try to keep thinking about conflicting ideas for longer periods and I find paradoxes deeply interesting. Much of my work is very interdisciplinary and I'm quite comfortable being an outsider in a new field of study.

News

I'm currently visiting Prof. Peter Dayan's lab at the Max Planck Institute for Biological Cybernetics to work on POMDP models of homeostasis.
Presented new upcoming work on optimal composition of multiple values, and resolving some mysteries of phasic dopamine at RLDM 2025.
Neural Associative Skill Memories has been accepted at the 5th International Workshop on Active Inference (IWAI 2024) as a Spotlight presentation.
My PhD's first paper, "Balancing safety and efficiency in human decision making", is now in eLife (eLife assessment: important work, convincing results).
Invited talk at 1st Oxford Health BRC Pain Conference (March 2024) on ’Virtual reality and Pain: From theory to practice’

Research vision

Key motivations in Neuro-AI research are twofold: utilising advancements from AI to generate useful hypotheses for neuroscience and leveraging insights from neuroscience to develop more effective algorithms for AI.

With these motivations in mind, here are some of the questions that motivate my recent work:

Continual Learning:
- How should intelligent agents continue to learn new skills and compose them optimally for reliable decision making?
- What sort of (temporal) abstractions in our world models best help us learn and plan effectively in new environments?
- What optimiser does the brain use? So as to enable efficient learning with local learning rules ...
Safe Natural Intelligence
- How do animals keep themselves safe? (e.g. safe exploration and self-preservation)
- How can we reverse-engineer our pain system and potentially fix it when it goes awry?
Safe Artificial Intelligence
- How can we make intelligent machines that act safely in human-like ways?
- How can we understand (unsafe) LLM misbehaviours through the lens of computational psychiatry and social neuroscience?
- How can we make AI trustworthy, interpretable, and aligned with human values in human-AI interactions?
AI and computational neuroscience for healthcare
- New technologies for better assessment and treatments to improve human health and well-being (with a recent focus on chronic pain)

Selected Publications

Safe exploration and neural reinforcement learning

Optimal composition of multiple values for dopamine-mediated efficient, safe and stable learning

RLDM conference (2025) and in preparation

Pranav Mahajan and Ben Seymour

Developed a theory that enables optimal composition of multiple value functions to drive reliable behaviour and efficient off-policy adaptation to dynamically changing rewards
Reconciled contradicting hypotheses about the computation of the tail of the striatum (threat prediction error and action prediction error), providing a unifying normative account

Balancing safety and efficiency in human decision making

eLife (2024) and COSYNE conference (2022)

Pranav Mahajan, Shuangyi Tong, Sang Wan Lee, Ben Seymour

Formulated a safety-efficiency trade-off in human reinforcement learning, proposing an uncertainty-based solution.
Designed and performed a Virtual Reality experiment to test the model and performed hierarchical Bayesian modeling of human behavior.

Paper

Webpage

Computational neuroscience and neuro-robotics

Neural Associative Skill Memories for safer robotics and modelling human sensorimotor repertoires

5th International Workshop on Active Inference (IWAI 2024) - Spotlight

and under review at Neural Computation.

Pranav Mahajan, Mufeng Tang, T. Ed Li, Ioannis Havoutis, Ben Seymour

Developed embodied generative models learned from demonstrations (in PyTorch) for safer robots (Franka Emika Panda arm), implementing fault detection and correction while explaining human sensorimotor experiment findings

Paper

Homeostasis After Injury: How Intertwined Inference and Control Underpin Post-Injury Pain and Behaviour

Under review at PLoS Computational Biology,

Pranav Mahajan, Peter Dayan, Ben Seymour

Theorised homeostasis after injury as planning under partial observability, explaining paradoxes and providing a novel perspective on existing literature.

Paper

Archived Projects

Hi, I'm Pranav.

News

Research vision

Selected Publications

Safe exploration and neural reinforcement learning

Optimal composition of multiple values for dopamine-mediated efficient, safe and stable learning

Balancing safety and efficiency in human decision making

Computational neuroscience and neuro-robotics

Neural Associative Skill Memories for safer robotics and modelling human sensorimotor repertoires

Homeostasis After Injury: How Intertwined Inference and Control Underpin Post-Injury Pain and Behaviour

Want to get in touch?

Email me at pranav[dot]mahajan[at]ndcn.ox.ac.uk