An Adaptive, Biologically-Inspired Framework for Identifying Salience in Data

In the modern world, people and machines are faced with ever-increasing amounts of data. Oftentimes, large data sets can be greatly reduced by identifying elements that are salient, within the context of a particular mission or goal. For example, in a military patrol video, hours of empty scenery may be removed to reveal short snippets where people or vehicles appear, and engage in behaviors that may indicate malicious intent. Generally, we require a methodology that can preprocess large, multimodal data sets by extracting salient information, which can then be used within decision support systems (or directly by humans) for further analysis. Intelligent Automation, Inc. and its partners propose to develop this methodology by taking inspiration from human perceptual processing. In particular, we shall leverage recent artificial neural network models that are able to control internal information flow, and thus give greater preference to more salient data elements. These models shall also have the ability to learn through reinforcement signals, which can be generated either manually or automatically to gauge performance. Thus, our methodology is highly general, and can adapt to different data modalities, domains and goals.