Loss Landscapes and Generalization in Neural Networks: Theory and Applications
| dc.contributor.advisor | Tran, Trac D | |
| dc.contributor.committeeMember | Patel, Vishal | |
| dc.contributor.committeeMember | Vidal, Rene | |
| dc.creator | Rangamani, Akshay | |
| dc.creator.orcid | 0000-0002-1961-0445 | |
| dc.date.accessioned | 2020-06-21T20:04:24Z | |
| dc.date.available | 2020-06-21T20:04:24Z | |
| dc.date.created | 2020-05 | |
| dc.date.issued | 2020-01-21 | |
| dc.date.submitted | May 2020 | |
| dc.date.updated | 2020-06-21T20:04:24Z | |
| dc.description.abstract | In the last decade or so, deep learning has revolutionized entire domains of machine learning. Neural networks have helped achieve significant improvements in computer vision, machine translation, speech recognition, etc. These powerful empirical demonstrations leave a wide gap between our current theoretical understanding of neural networks and their practical performance. The theoretical questions in deep learning can be put under three broad but inter-related themes: 1) Architecture/Representation, 2) Optimization, and 3) Generalization. In this dissertation, we study the landscapes of different deep learning problems to answer questions in the above themes. First, in order to understand what representations can be learned by neural networks, we study simple Autoencoder networks with one hidden layer of rectified linear units. We connect autoencoders to the well-known problem in signal processing of Sparse Coding. We show that the squared reconstruction error loss function has a critical point at the ground truth dictionary under an appropriate generative model. Next, we turn our attention to a problem at the intersection of optimization and generalization. Training deep networks through empirical risk minimization is a non-convex problem with many local minima in the loss landscape. A number of empirical studies have observed that “flat minima” for neural networks tend to generalize better than sharper minima. However, quantifying the flatness or sharpness of minima has been an issue due to possible rescaling in neural networks with positively homogenous activations. We use ideas from Riemannian geometry to define a new measure of flatness that is invariant to rescaling. We test the hypothesis that flatter minima generalize better through a number of different experiments on deep networks. Finally we apply deep networks to computer vision problems with com- pressed measurements of natural images and videos. We conduct experiments to characterize the situations in which these networks fail, and those in which they succeed. We train deep networks to perform object detection and classification directly on these compressive measurements of images, without trying to reconstruct the scene first. These experiments are conducted on public datasets as well as datasets specific to a sponsor of our research. | |
| dc.format.mimetype | application/pdf | |
| dc.identifier.uri | http://jhir.library.jhu.edu/handle/1774.2/62471 | |
| dc.language.iso | en_US | |
| dc.publisher | Johns Hopkins University | |
| dc.publisher.country | USA | |
| dc.subject | Signal Processing | |
| dc.subject | Machine Learning | |
| dc.subject | Deep Learning | |
| dc.subject | Deep Neural Networks | |
| dc.subject | Autoencoders | |
| dc.subject | Representation Learning | |
| dc.subject | Dictionary Learning | |
| dc.subject | Flat Minima | |
| dc.subject | Riemannian Geometry | |
| dc.title | Loss Landscapes and Generalization in Neural Networks: Theory and Applications | |
| dc.type | Thesis | |
| dc.type.material | text | |
| thesis.degree.department | Electrical and Computer Engineering | |
| thesis.degree.discipline | Electrical Engineering | |
| thesis.degree.grantor | Johns Hopkins University | |
| thesis.degree.grantor | Whiting School of Engineering | |
| thesis.degree.level | Doctoral | |
| thesis.degree.name | Ph.D. |