My research lies at the intersection of neural network interpretability, computational neuroscience, and signal processing. I am particularly interested in understanding the internal representations of artificial neural networks and biological brains, by developing tools for interpretability. In my recent work, I have shown that interpretability tools need to be designed using suitable inductive biases about the geometric structure of concepts in deep neural network activations, failing which they will not be able to capture those concepts. With my collaborators, I have demonstrated the importance of inductive biases by constructing suitable tools such as:
SpaDE (Sparsemax Distance Encoder) to capture concepts that are nonlinearly separable and heterogeneous (Hindupur et al, NeurIPS 2025),
Temporal Feature Analyser (TFA) to capture concepts associated with temporal structure, such as new concepts learnt in-context, event boundaries in narrations, parsing of garden-path sentences (Lubana, Rager, Hindupur et al, arXiv 2025)
SparKer (Sparse ensemble of local Kernels) to identify statistical anomalies in data across sciences (particle physics) and ML (word embedding models) (Grosso et al, arXiv 2025)
"Sparse, self-organizing ensembles of local kernels detect rare statistical anomalies" [paper link]
Gaia Grosso, Sai Sumedh R. Hindupur, Thomas Fel, Samuel Bright-Thonney, Philip Harris, Demba Ba
Preprint on arXiv (2025)
"Priors in Time: Missing Inductive Biases for Language Model Interpretability" [paper link]
Ekdeep Singh Lubana*, Can Rager*, Sai Sumedh R Hindupur*, et al
Preprint on arXiv (2025)
"Projecting Assumptions: The Duality between Sparse Autoencoders and Concept Geometry" [paper link]
Sai Sumedh R. Hindupur*, Ekdeep Singh Lubana*, Thomas Fel* and Demba Ba
Published in Advances in Neural Information Processing Systems (NeurIPS) 2025
"Population Coding under the Scale-Invariance of High-Dimensional Noise" [paper link]
S. Amin Moosavi, Sai Sumedh R. Hindupur, Hideaki Shimazaki
Preprint on bioRxiv (2024)
"Online Parameter Estimation in Partially Observed Markov Decision Processes" [paper link]
Sai Sumedh R. Hindupur, Vivek S Borkar
Published at the 59th Annual Allerton Conference on Communication, Control, and Computing (2023)
*denotes equal contribution (co-first authors)
Posters and Presentations
Priors in Time: A Generative View of SAEs for Sequential Representations [link]
at the 2nd New England Mechanistic Interpretability Workshop (NEMI), Northeastern University, Boston, MA (August 22, 2025)
Projecting Assumptions: The Duality between Sparse Autoencoders and Concept Geometry [link]Ā
Methods and Opportunities at Small Scale (MOSS) Workshop at ICML 2025, Vancouver, Canada (July 19, 2025)
Invited Speaker at Caltech NeuroAI Journal Club (virtually on Zoom) (May 28, 2025)
New England NLP Workshop (NENLP), Yale University, New Haven, CT (April 11, 2025)
2nd Spring into Science by the Kempner Institute at Harvard University, Loft on Two, Boston, MA (March 26, 2025)
SPrInt: Sparsemax Prototypes for Interpretability [link]
atĀ theĀ 1st New England Mechanistic Interpretability Workshop (NEMI), Northeastern University, Boston, MA (August 19, 2024)
In my undergrad+masters degree at IIT Bombay, I worked on multiple research projects:
Online parameter estimation in POMDPs [paper] (with Prof. Vivek Borkar, IIT Bombay)
This work introduces a novel methodology for real-time parameter estimation in POMDPs using maximum likelihood estimation integrated with extremum-seeking. The framework enables online learning for sequential decision-making processes with partial observations and is supported by theoretical analyses and extensive numerical validations.
Neural population coding in the mouse visual cortex [paper](in collaboration with Dr. Amin Moosavi, UCLA and Prof. Hideaki Shimazaki, Kyoto University),
This study explores the role of scale-invariant noise structures in high-dimensional neural activity and their implications for population coding. Using large-scale neural recordings from the mouse visual cortex, we identify conditions under which neural populations achieve unbounded information capacity, offering insights into the scalability of neural coding systems.
Spiking neural network architectures [report] (with Prof. Alice Parker, University of Southern California),
We simulated a biologically inspired spiking neural network to tackle catastrophic forgetting in sequential learning tasks. By leveraging mechanisms like dopamine-driven reward learning and synaptic plasticity, the NeuRoBot successfully retained previously learned skills while acquiring new ones. This work bridges neuroscience insights with adaptive AI design.
MEG-based signal analysis of neural activity [report] (with Prof. Ole Jensen and Dr. Yali Pan, University of Birmingham).
We investigated interactions between brain oscillations during reading by analyzing MEG data. Using rapid frequency tagging and statistical tests, we studied subtle interactions between alpha oscillations and high-frequency activity in the visual cortex.
