Presenter
Gerald Pao
Dr. Gerald Pao began his scientific career studying protein evolution at UC San Diego, working with Milton Saier, Joseph Kraut, and others. He then pursued epigenetics and viral vector development at the Salk Institute during his PhD and postdoc with Inder Verma. His work on stem cells and regeneration led him to study axolotls with David Gardiner and Tony Hunter. Shifting focus, he trained in nonlinear dynamics and data science at Scripps Institution of Oceanography with George Sugihara. At the Salk Institute, he applied these methods to systems neuroscience and biology. He was also a visiting scientist at Japanโs AIST, optimizing computational methods for Big Data on the ABCI supercomputer. Additionally, he developed cephalopod reflectin proteins for in vivo optical manipulation. Before joining OIST, Dr. Pao led data science and gene therapy initiatives at Vertex Pharmaceuticals.
Abstract:
Quantitative science has long been dominated by physics, which aims to express relationships among natural variables through equations. These equations ideally require linearity, decomposability, and Gaussian noise for analytical solutions. However, such assumptions break down for highly nonlinear systems, where traditional physics-based approaches have struggled. In contrast, Deep Learning has shown surprising success in capturing nonlinear mappingsโbut at the cost of explainability due to its โblack boxโ nature. Neural activity is profoundly nonlinear. To address this with interpretability, we introduce a framework based on the generalized Takens theorem from dynamical systems theory. It enables data-driven embeddings of time series on low-dimensional manifolds that map neural activity to behavior. This allows for prediction, causal inference at scale, simulation, and guarantees of explainability and testability. We demonstrate several use cases in animal and human systems neuroscience that effectively simulate neural and behavioral dynamicsโakin to single-behavior brain downloads.
