| Sol No.: |
Navy SBIR FY2012.2 |
| Topic No.: |
N122-138 |
| Topic Title: |
Interactive Generative Manifold Learning |
| Proposal No.: |
N122-138-0014 |
| Firm: |
ObjectVideo
11600 Sunrise Valley Drive
Suite # 210
Reston, Virginia 20191 |
| Contact: |
Ping Wang |
| Phone: |
(703) 654-9352 |
| Web Site: |
www.objectvideo.com |
| Abstract: |
Exploratory data analysis is the foremost step in selecting appropriate statistical learning algorithms specialized to a dataset. We have proposed a generative framework for visualizing high dimensional data as low-dimensional manifold embedded in a high dimensional space. The method allows user to conveniently explore the space using fewer dimensions while still capturing the principal modes of variations of the high dimensional data. Specifically, we employ Gaussian Process Latent Variable Model (GPLVM) and Spectral Latent Variable Model (SLVM) to learn low-dimensional representations of the data. Probabilistic mappings between the embeddings and the original space facilitate efficient interpolation in the latent space as well as fast visualization of the interpolated latent points in the original space. To allow the user to span the manifold in an intuitive manner, we develop supervised and semi-supervised tools to relate the latent space to the meaningful feature space. These enable computation of principle direction for each point in the latent space to allow the user to traverse in a meaningful way. Further, we have proposed a principled approach to extrapolate the latent space by predicting the manifold structure in regions lying outside the existing domain of the data. |
| Benefits: |
The technology developed as a part of this effort is a core component in all statistical learning frameworks. The anticipated applications of this work include:
a) Generation of novel synthetic data from the learned models - The tool can be used to synthesize novel data by randomly moving in the learned low-dimensional space and back projecting the points to original space;
b) Object classification / recognition using selective meaning features - The proposed approach can be directly applied to classify objects by projecting to low-dimensional manifold and comparing them along semantically meaningful dimensions;
c) Selecting appropriate machine learning tools specialized to the data set - The tool can be used to compare whether linear or non-linear methods are applicable to the dataset. It may also be used to compare different dimensionality reduction techniques for a given dataset;
d) Clean the noisy data by projecting onto the embedded space and back-projecting the latent points to original space.
|