Luca Herranz-Celotti

He holds a B.Sc. in Physics and an M.Sc. in Biophysics from the Autonomous University of Madrid, and a Ph.D. in Machine Learning and Computational Neuroscience from the University of Sherbrooke (Québec, Canada), where he specialized in Gradient Stability. His doctoral work demonstrated the effectiveness of Gradient Stability as a framework for understanding biologically plausible spiking neurons. He also clarified a long-standing misinterpretation of gradient explosion in deep recurrent networks, showing that an additive exponential effect had been inaccurately characterized. During his Ph.D., he proposed pretraining to stability as a strategy for improving the generalization performance of a wide variety of deep recurrent architectures.

In collaboration with MILA, he contributed to the HoME, the first dataset for multimodal embodied learning. He also designed U-BESD, a neural network that helps people with hearing impairments isolate the sounds they want to focus on. In applied NLP, he has worked on privacy-preserving information retrieval systems robust to misinformation by fine-tuning large language models to solve faulty-logic reasoning tasks.

His current research focuses on State-Space Models for language understanding and generation, as well as multi-agent neurosymbolic systems aimed at overcoming the limitations of LLMs in non-monotonic reasoning.