Hello, I'm Farooq.
I am a machine learning researcher. I study what deep models actually learn when their supervision is imperfect, and how to tell from inside the model: from its representations, its spectra and its weights, and not only from its outputs.
In September 2026 I defended my PhD in Data Science at Sapienza University of Rome, advised by Prof. Fabrizio Silvestri, with the thesis Robust Representation Learning under Imperfect Supervision. From November 2026 I am a postdoctoral researcher at the Free University of Bozen-Bolzano, where I work with Prof. Raffaella Bernardi on LLADIGA (Learning Language with Dialogue Games): evaluating the language-use capabilities of large language models and analysing what learning in interaction changes inside them.
What I work on
Vision-language models. I look for the instability that a correct answer can hide (EMNLP 2026), and at how the wording of a prompt changes robustness.
Learning with noisy labels. I design losses that limit how much a network memorises wrong labels, on images and on graphs (NCOD, GCOD, and a public competition at IJCNN 2025).
Repairing trained models. I work on removing the effect of corrupted training data from a model when the original data are no longer available (CUTS, ECML PKDD 2026).
Generalisation on graphs. I work on graph classifiers that generalise when the test graphs come from a different distribution than the training graphs (PISA, LoG 2026).
Before the PhD I worked for four years as a data scientist at General Electric in Bangalore, and I hold an MSc in Artificial Intelligence and Robotics from Sapienza. During the PhD I was a visiting researcher at the University of Edinburgh, hosted by Dr Pasquale Minervini. I have also contributed to work on audio deepfake detection.