In situ learning for underwater object recognition Navy SBIR FY2009.1 Sol No.: Navy SBIR FY2009.1 Topic No.: N091-066 Topic Title: In situ learning for underwater object recognition Proposal No.: N091-066-0370 Firm: Signal Innovations Group, Inc. 1009 Slater Rd. Suite 200 Durham, North Carolina 27703 Contact: Patrick Rabenold Phone: (919) 323-3452 Web Site: www.siginnovations.com Abstract: We propose a principled in situ learning framework that is appropriate for a Bayesian classifier implemented with semi-supervised and multi-task learning. We will investigate several different forms of in situ learning, and will perform testing on measured data to help define which is most appropriate for Navy sensing missions. In addition, we will develop new techniques for feature adaptivity and selection, to tune the features to the particular targets and clutter in the environment under test. Benefits: The proposed research has the opportunity to significantly advance the manner in which the Navy performs underwater sensing. The proposed algorithms will yield a high level of adaptivity, allowing for changing environmental, target and clutter conditions. This will be performed using a new class of in situ learning algorithms, that allow the classifiers to adapt to sensing conditions. The proposed algorithms are appropriate for integration within semi-supervised and multi-task learning, these exploiting a significant level of context. In the proposed research we will not explicitly develop semi-supervised and multi-task algorithms, as these are being developed by SIG under separate support. However, we will leverage that research, and the proposed in situ learning algorithms will be integrated into such. The proposed Phase I research will make the following specific contributions: (1) The proposed algorithms will take into account real-world sensing issues, such as the potential for noisy data/labels, as well as the cost of label/data acquisition. (2) A new Bayesian Elastic Net algorithm is proposed, which will provide a level of accuracy in feature selection that has previously been unavailable. In a Phase II effort, these new approaches will be integrated into the semi-supervised multi-task learning algorithms, providing a unique framework for sensing adaptivity and in situ learning. The potential commercial applications of the research are as follows. The availability of high spatial and spectral resolution satellite imagery has created a growth industry in land-use assessment, optimized natural resource extraction, habitat analysis, precision agriculture, and urban planning/infrastructure analysis. However, the variability of sensing parameters (i.e., cloud-cover, seasonal variations, incident and scattering angles relative to time of day and sensor orientation) can confound classification performance between training vs. operational imagery. Moreover, the amount of labeled training data is small relative to the volume of spatial/spectral imagery generated by modern commercial satellites such as Ikonosand LandSat 7. The algorithms developed under this effort could accommodate such high data volumes and adapt to sensor variability for improved image scene characterization. Such methods could be useful to the Bureau of Land management (BLM), the United States Geologic Service (USGS), the United States Department of Agriculture (USDA), and the intelligence community for the analysis of commercial, and non-commercial, remote sensing data. Another key application of the proposed research is in the area of bioinformatics. For example, when dealing with gene-expression data one typically has large quantities of highdimensional gene-expression feature vectors, with few labeled training examples. The labeling of this data is very expensive and labor intensive, requiring the attention of experts. The in situ learning algorithms may be used to provide feedback to medical researchers, as to which feature vectors are most informative if they could be labeled (for example, by a human expert, or via conventional diagnoses). Moreover, the large quantities of unlabeled data make this a good application for semisupervised algorithms, in which all available data are exploited when performing algorithm design (both labeled and unlabeled data). Initial forms of these algorithms were developed recently by Prof. Carin at Duke, and several of his recent PhD graduates have been employed by medical companies such as SIEMENS, GE and Guidant. It is therefore anticipated that such companies represent a wide market for the class of software tools we propose to develop under this effort. Return