Friday, May 27, 2022 • 3:30–4:45 pm • Light Hall, Room 208
Organizer: Chunming Zhang, University of Wisconsin–Madison
Chair: Samuel Davenport, University of California San Diego
On network modularity statistics in connectomics and schizophrenia
Joshua Cape, University of Pittsburgh
Anirban Mitra, Indiana University
Konasale Prasad, University of Pittsburgh
Modularity-based methods for structure and community discovery remain popular in the network neuroscience literature and enjoy a history of yielding meaningful neurobiological findings. All the while, the full potential of these methods remains limited in part by an absence of uncertainty quantification guarantees for use in downstream statistical inference. Here, we pursue this direction by revisiting the classical notion of modularity maximization in the analysis of adjacency and correlation matrices. We begin by considering certain latent space network models wherein high-dimensional matrix spectral properties can be precisely analyzed. We further propose and argue for the potential usefulness of several new, non-classical modularity-type network statistics. Our findings are applied to an analysis of dMRI and fMRI data in the study of schizophrenia.
Robustness (or not) of deep learning methods applied to medical images
Alan B. McMillan, University of Wisconsin
Deep learning methods are extensively applied to various medical image reconstruction, processing, and analysis pipelines. However, deep learning is known to be particularly brittle when exposed to adversarial inputs—either of intentional or unintentional origin. In this talk, we discuss strategies to assess and improve the robustness of deep learning methods for medical images.
Imitation learning of spatio-temporal point processes
Yao Xie, Georgia Institute of Technology
Shixiang Zhu, Georgia Institute of Technology
Shuang Li, Harvard University
Zhigang Peng, Georgia Institute of Technology
We present a novel Neural Embedding Spatio-Temporal (NEST) point process model for spatio-temporal discrete event data and develop an efficient imitation learning (a type of reinforcement learning) based approach for model fitting. Despite the rapid development of one-dimensional temporal point processes for discrete event data, the study of spatial-temporal aspects of such data is relatively scarce. Our model captures complex spatio-temporal dependence between discrete events by carefully design a mixture of heterogeneous Gaussian diffusion kernels, whose parameters are parameterized by neural networks. This new kernel is the key that our model can capture intricate spatial dependence patterns and yet still lead to interpretable results as we examine maps of Gaussian diffusion kernel parameters. The imitation learning model fitting for the NEST is more robust than the maximum likelihood estimate. It directly measures the divergence between the empirical distributions between the training data and the model-generated data. Moreover, our imitation learning-based approach enjoys computational efficiency due to the explicit characterization of the reward function related to the likelihood function; furthermore, the likelihood function under our model enjoys tractable expression due to Gaussian kernel parameterization. Experiments based on real data show our method's good performance relative to the state-of-the-art and the good interpretability of NEST's result.