Learning Robust Classifiers from Small Data using Generative Models
Abstract
Deep neural networks have achieved state-of-the-art results on a range of supervised tasks, such as image classification, speech recognition, and machine translation. However, these successes were often enabled by large quantities of labeled data, e.g., the ImageNet database. This presents a critical challenge if the availability of labels for the target task is limited, as in several military applications. Synthetic data generation, either through simulators or statistical generative models, is a promising approach to overcome this challenge. Unfortunately, synthetic data is always imperfect -- a problem known as domain shift. This often leads to limited practical benefits compared to labeled data collected in the real world. The goal of this proposal is to bridge this gap, and develop more effective learning techniques in the presence of domain shifts. Specifically, we aim to develop algorithms to 1) enable learning with less labeled data by 2) increasing robustness to domain shifts. Focusing on image classification, we propose to develop new approaches to: 1. Detect and quantifying domain shifts, so as to enable transfer of knowledge from one domain to another 2. Perform domain-adapted semi-supervised learning to enable learning with limited labels in the presence of domain shifts. 3. Learn in the presence of extreme domain shifts, such as those caused by incorrect data 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 image recognition and analysis benchmarks.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 25, 2021
- Source ID
- W911NF2110125
Entities
People
- Stefano Ermon
Organizations
- Army Contracting Command
- Stanford University
- United States Army