I am a PhD student working on theoretical machine learning, with a focus on composition, memory and pruning.
Based in Nice, France, I am affiliated with COATI, 3IA Côte d’Azur, Inria, I3S, CNRS, and Université Côte d’Azur. Working under the supervision of Prof. Emanuele Natale, my research studies the mathematical principles behind model composition, sequence memory, and neural network pruning.
Prior to this, I completed my BS-MS in Physics at IISER Kolkata, giving me a strong foundation in the physical principles that often govern complex systems like Neural Networks.
Compositionality, Memory, and Network Structure
A central part of my research studies the Strong Lottery Ticket Hypothesis (SLTH), which asks whether sparse subnetworks hidden inside large randomly initialized models can already exhibit strong performance before training. My earlier work developed theoretical frameworks connecting pruning and quantization in these regimes.
More recently, I have been exploring how neural systems combine specialized behaviors without interference. One line of work develops a principled framework for composing autoregressive models, with guarantees on stability and compositional generalization.
Another line investigates asymmetric Hopfield networks and sequence memory, showing classical asymmetric Hopfield networks can support extremely large number of stable limit cycles, each with a very large period.
Team COATI. Advisor: Emanuele Natale.
Thesis: "Phase transition in Artificial Neural Networks".
Teaching Assistant for Mathematical Methods I.
