Virtual Talk: Topology-Driven Machine Learning - Chao Chen via Zoom
Abstract: Modern machine learning is blessed and cursed with highly complex models. While deep neural networks have proven powerful in practice, their deployment in real-world applications faces challenges such as explanability and robustness. One crucial reason is the lack of sufficient means to track/control the underlying structure of the model/data. In this talk, we present examples in which advanced structural prior can be encoded into learning through the mathematical language of topology. We will explain how global topological structures, such as connected components and handles/loops, can be efficiently and robustly extracted from the data, as advanced features and even differentiable loss functions. In biomedical applications, e.g., neuron and cell image analysis, we proposed a novel topological loss function that enforces the neural network to learn the topology from the data. In machine learning, we used topology of data represented at different layers to improve the robustness of a neural network against noise and potentially against malicious attacks. We will also show how advanced topological and geometric information can help graph neural networks to learn from graph-structured data. The take-home message is that with the help of deep neural networks as advanced data-denoising tools, classic topological and geometric information can be even more reliably computed and in turn be used to improve the transparency and robustness of the complex learning models.

Bio: Chao Chen is an Assistant Professor from the Department of Biomedical Informatics. He is also affiliated with the Departments of Computer Science and Applied Mathematics & Statistics. Originally trained in computational geometry/topology, he is particularly interested in developing new methods based on principled topological and geometric information, and uses these methods to address challenges in machine learning and in biomedical data analytics. More information can be found at: https://bmi.stonybrookmedicine.edu/people/chao_chen.

For more Zoom details, please email: events [at] cs.stonybrook.eduid="ow863" __is_owner="true".