A Hypothesis-based Dictionary Learning Model

Recent research in machine learning and signal processing has focused on modeling data with linear combinations of elements from learned dictionaries. We propose a hypothesis-based framework for classification tasks that classifies signals according to their structured dictionary representations. The framework considers that signal representations display different structures depending on their label information, and it makes hypotheses to analyze how likely a signal belongs to a particular class. The likelihood of belonging to a class is revealed by the patterns of representations on the dictionary, which allows for distinguishing signals from different categories. Additionally, our proposed framework is computationally more efficient than traditional sparse dictionaries. We validate the efficiency and performance of our framework on standard datasets used for evaluating dictionary learning models.