A1410
Title: Induced likelihood-based methods for soft classification of u-processes arising in few shot learning problems
Authors: Christopher Saunders - South Dakota State Univerisity (United States) [presenting]
Janean Hanka - South Dakota State University (United States)
Danica Ommen - Iowa State University (United States)
Semhar Michael - South Dakota State University (United States)
Abstract: Few-shot learning problems are the most common class of pattern recognition tasks commonly encountered in forensic source identification. The data resources for this problem are commonly structured in a way that involves a large number (denoted as $C$) of classes (typically referred to as sources), each with two to five exemplars (typically referred to as control samples or knowns) per class of interest. Typically, a metric for measuring the dissimilarity between a pair of objects is constructed or learned from earlier studies. This learned metric is used to make predictions as a pseudo-metric for a $k-$nearest-neighbor classifier or as a metric for a kernel density estimator; neither of which leads to a probabilistic interpretation of the final output. Furthermore, the limited sample sizes within each class limit the ability to use a brute force method such as kernel density. To work around these structural constraints, a clustering-based method is used for the empirical distribution of pairwise scores within each class, which allows the pooling within distributions of scores across sources in a statistically rigorous manner to make stable probabilistic statements.
