(Imperial College

The global COVID19 pandemic has highlighted the lethality and morbidity associated with infectious respiratory diseases. These diseases can lead to devastating syndrome known as the acute respiratory distress syndrome (ARDS) where bacterial/viral infections cause excessive lung inflammation, pulmonary edema, and severe hypoxemia (shortness of breath). Although ARDS patients require artificial mechanical ventilation, the complex biofluid and biomechanical forces generated by the ventilator exacerbates lung injury leading to high mortality. My group has used mathematical and computational modeling to both characterize the complex mechanics of lung injury during ventilation and to identify novel ways to prevent injury at the cellular level. We have used in-vitro and in-vivo studies to validate our mathematical predictions and have used engineering tools to understand the biological consequences of the mechanical forces generated during ventilation. In this talk I will specifically describe how our mathematical/computational approach has led to novel cytoskeletal based therapies and how coupling mathematics and molecular biology has led to the discovery of a gene regulatory mechanisms that can minimize ventilation induced lung injury. I will also describe how we are currently using nanotechnology and gene/drug delivery systems to enhance the lung’s native regulatory responses and thereby prevent lung injury during ARDS.

Training deep learning models involves searching for a good model over the space of possible architectures and their parameters. Discovering models that exhibit robust generalization to unseen data and tasks is of paramount for accurate and reliable machine learning. Generalization, a hallmark of model efficacy, is conventionally gauged by a model's performance on data beyond its training set. Yet, the reliance on vast training datasets raises a pivotal question: how can deep learning models transcend the notorious hurdle of 'memorization' to generalize effectively? Is it feasible to assess and guarantee the generalization prowess of deep neural networks in advance of empirical testing, and notably, without any recourse to test data? This inquiry is not merely theoretical; it underpins the practical utility of deep learning across myriad applications. In this talk, I will show that scrutinizing the training dynamics of neural networks through the lens of topology, specifically using 'persistent-homology dimension', leads to novel bounds on the generalization gap and can help demystifying the inner workings of neural networks. Our work bridges deep learning with the abstract realms of topology and learning theory, while relating to information theory through compression.