Uncertainty-aware Learning with Generative Models
Abstract
Deep neural networks have achieved state-of-the-art results on a range of supervised tasks, suchas image classification, speech recognition, machine translation, and game playing. However, achieving high-performance with deep learning on a new task often requireslarge quantities of labeled data, e.g., the ImageNet database. Additionally, deep learning solutions are vulnerable to adversarial attacks and often experience difficulties in extrapolating beyond the training data, likely due to their inability to reason about objects, their relationships, and physics. Furthermore, for the same reasons changes in operating conditions after deployment often require expensive re-training, rendering deep learning solutions difficult to deploy in high-stakes military applications.To overcomethese limitations, we aim to develop algorithms to 1) enable learning with less labeled data and 2) increase robustness and uncertainty quantification capabilities, i.e., develop models that know what they do not know. To achieve these next-generation AI capabilities, we need to go beyond traditional discriminative models (which can only map inputs to the corresponding label) and embrace fully generative models that can also reason about the inputs they receive. This will enable learning from large quantities of unlabeleddata (i.e., inputs without the corresponding ground-truth outputs) through principled unsupervised learning algorithms and improveduncertainty quantification capabilities (e.g, by detecting anomalous or out-of-distribution inputs). The research plan consists of:(1) a theoretical component involving the design of new probabilistic models, learning objectives, and analysis of their theoretical properties, and (2) practical experimental evaluation on standard benchmarks.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 06, 2023
- Source ID
- N000142312159
Entities
People
- Stefano Ermon
Organizations
- Office of Naval Research
- Stanford University
- United States Navy