Karim Haroun

Hi everyone! Welcome to my webpage. My name is Karim Haroun, I am a research scientist in Machine Learning, where my main research interest revolves around neural network compression for more frugal ML algorithms. I am also a strong believer and advocator of openscience that is beneficial for society. When bored, I happen to pick my classical guitar and forget about everything.

From March 2022 to May 2025, I was a PhD student affiliated to the Embedded Artificial Intelligence Lab (LIAE) from CEA-LIST, and the SPARKS team at the Laboratory of Computer Science, Signals and Systems of Sophia Antipolis (I3S), Université Côte d'Azur. In my thesis, I worked on dynamic neural network compression, where the computational graph is adapted to input hardness to achieve input-adaptive computation. I focused on token merging strategies for Transformers that vary the sequence length by leveraging token redundancy.

My current research focuses on theoretical aspects of neural networks, including more principled neural network compression using informative measures, modularity in deep learning, neural architecture growth, and generative models. I am also exploring applications in dense computer vision tasks and natural language processing. Last, I was recently granted a 5k euros funding from the Centre interdisciplinaire des innovations écologiques et sociales (CITIES), Université Paris 8, to support new research on fairness and biases of Large Language Models (LLMs).