Efficient, Robust and Reliable Neural Networks for Multimodal and Synthetic Data: A Sparse Representation Perspective
Abstract
Modern machine learning models, such as Deep Neural Networks (DNNs), where the number of parameters is quite large, have been extremely successful in supervised learning and generative modeling tasks. However, despite their prevalence and the surge of interest, there are major gaps in our understanding of the fundamental behavior of DNN models. Understanding the fundamental mechanism behind their generalization properties remains a challenge. Training and tuning DNN based models is tremendously challenging, especially with multimodal data sources. The most significant obstacle is the high dimensional non-convex optimization landscape as a result of end-to-end training. Moreover, most data intensive tasks relevant to army s mission involve multimodal and heterogeneous data sources coming from various sensors and systems of interconnected objects, which further exacerbate these issues. DNN models in multimodal data settings are extremely difficult to tune, troubleshoot and trust. Furthermore, recent work has shown that neural network based learning is quite brittle where small adversarial changes can lead to incorrect results with very high confidence. On the other hand, generative models also suffer from the black-box and complex nature of DNNs. Training trustworthy, robust and interpretable generative models from multimodal and heterogeneous data sources remains an open problem. This research proposal brings together a diverse body of mathematical theories and ideas to address key questions including the following: (1) How to establish theoretical foundations of over-parameterized models such as deep neural networks? (2) How to design optimal learning architectures for multimodal data and provide theoretical guarantees? (3) Can the sample efficiency and training complexity of neural models analyzed and improved? (4) Can real data be replaced by generative models for training and testing, and how this can be quantified? (5) How to come up with testable, quantifiable metrics for the quality, robustness and trustworthiness of multimodal neural network models? This project promises to develop a novel framework for trustworthy, transparent and robust machine learning techniques, including multimodal neural networks and generative models. In particular, part of the proposed theory leverages a revolutionary hidden convex regularization theory for neural networks. The proposed approach will demystify current deep neural networks, and potentially replace it with efficient, transparent and trustworthy models with guarantees. Moreover, the proposed research will establish previously unexplored and unexpected connections to sparse representations, which will give rise to new theoretical and algorithmic developments. The research results will enable significantly better performing, more sample efficient, robust and safe neural models which are easy to use in practical problems.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 25, 2021
- Source ID
- W911NF2110242
Entities
People
- Mert Pilanci
Organizations
- Army Contracting Command
- Stanford University
- United States Army