Hyperparameter Optimization for Object and Event Detection from Patterns Using Deep Learning
Abstract
Object and event recognition from streaming sensor data is an important problem with applicability to a variety of domains. These include robotics and automation, video semantic understanding, and situational awareness from surveillance. Because of the importance of this object recognition problem to a broad range of applications, researchers have produced a number of solutions. Recently, deep learning neural networks have shown particular success for this problem. These deep learning methods have essentially replaced the conventional approaches to the problem that used manual, expert knowledge to engineer features which were then input to classi ers of various types. Instead the deep learning neural nets accept \raw" (normalized) inputs and learn relevant features as they learn to correctly classify. A problem with this approach is that the architectures of these deep learning neural networks have a large number of hyperparameters that must be set. While a few researchers have developed optimization approaches for these hyperparameters, their comparative e ectiveness has not been systematically studied for the important object recognition problem. The work proposed here addresses this shortcoming by conducting a comparative analysis of hyperparameter optimization techniques for deep learning neural networks applied to object recognition from streaming spatial, temporal data. The results from this work will provide an excellent foundation for signi cant performance improvements using the deep learning networks across a wide range of applications or, if the results are negative, show that deep learning neural networks can perform well with simple grid search or manual hyperparameter selection
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 06, 2020
- Source ID
- N002441910005
Entities
People
- Donald E. Brown
Organizations
- United States Navy
- University of Virginia