Interpretable and Generalizable Deep Neural Nets

Interpretable and Generalizable Deep Neural Nets

PI: Hongcheng Liu, Co-PI: Panos Pardalos
Award Period: 05/01/2019-06/30/2020
Abstract

Neural network (NN) is the pillar machine learning model for the realization of modern artificial intelligence. Despite the frequent advancement in the NN-related algorithms, models, and applications, the development of their theoretical underpinnings are lagging behand. For most existing theoretical generalization analysis, the number of samples required to ensure proper out-of-sample performance is stipulated to grow rapidly in polynomial of the number of fitting parameters and, thus, the depth of a NN. This is inconsistent with the practical observations, as modern neural network models are commonly over-parameterized. Furthermore, most existing NN models are hardly interpretable due to the intrinsic nonlinearity and nonconvexity. Through this project, we will theoretically analyze both the generalizability and the interpretability for a neural network in a general setting. The new bound will ensure the generalization performance of a NN to be insensitive to the increase of fitting parameters. Furthermore, regularization schemes will be incorporated into the training of a NN to make available a likelihood ratio-based statistical test, allowing the NN to be better interpretable. This project will lay foundation for the future research on artificial intelligence and global optimization by providing a machinery to comprehend a class of machine learning models. It will also lead to further development of a project relevant to one of NSF’s 10 big ideas: Harnessing Data for the 21st Century. At completion of this project, joint-author papers and proposals to NSF will be submitted.

Publications
  1. Seonho Park, Seung Hyun Jung, Panos Pardalos. Combining Stochastic Adaptive Cubic Regularization with Negative Curvature for Nonconvex Optimization, Journal of Optimization Theory and Application 184(3): 953-971, (March 2020).
  2. Stamatios-Aggelos N. Alexandropoulos, Panos M. Pardalos, Michael N. Vrahatis:
    Dynamic search trajectory methods for global optimization. Ann. Math. Artif. Intell. 88(1): 3-37 (2020)
  3. Hongcheng Liu, Yinyu Ye. (Under Review).High-Dimensional Learning under Approximate Sparsity: Towards a Unified Framework for Nonsmooth Learning and Regularized Neural Networks. Major revision by Operations Research. (2020)
  4. Qingchao Zhang, Yunmei Chen, Hongcheng Liu, Xiaojing Ye. (Under Review). A Novel Learnable Gradient Descent Type Algorithm for Non-convex Non-smooth Inverse Problems. Under review by ECCV.
  5. Bijan Taslimi, Yuanbo Wang, Hongcheng Liu, Panos Pardalos. (In Progress). A Lasserre Hierarchy-based global optimization scheme for training neural networks. Working paper.
Proposals
  1. Sponsor: Air Force Office of Scientific Research
    Title: Transforming Training Paradigms for Artificial Intelligence (White Paper)
    PI: Hongcheng Liu, Co-PI: Panos Pardalos
  2. Sponsor: Office of Naval Research
    Title: A Novel Interior-Point Trust-Region Paradigm for Optimization under Conic Constraints (White Paper)
    PI: Hongcheng Liu
  3. Sponsor: UF Office of Research
    Title: OR: DRPD-ROF2020: Generalizable Deep Neural Nets and Deep Hypothesis Testing for Biomarker        Identification and Disease Diagnosis
    PI: Hongcheng Liu, Co-PI: Panos Pardalos
  4. Sponsor: National Science Foundation
    Title: AF: Small: Data-Driven Optimization Under Partial Knowledge and Data Insufficiency
    PI: Hongcheng Liu, Co-PI: Yunmei Chen