Top-Down Selection in Convolutional Neural Networks

Feedforward information processing fills the role of hierarchical feature encoding, transformation, reduction, and abstraction in a bottom-up manner. This paradigm of information processing is sufficient for task requirements that are satisfied in the one-shot rapid traversal of sensory information through the visual hierarchy. However, some tasks demand higher-order information processing using short-term recurrent, long-range feedback, or other processes. The predictive, corrective, and modulatory information processing in top-down fashion complement the feedforward pass to fulfill many complex task requirements. Convolutional neural networks have recently been successful in addressing some aspects of the feedforward processing. However, the role of top-down processing in such models has not yet been fully understood. We propose a top-down selection framework for convolutional neural networks to address the selective and modulatory nature of top-down processing in vision systems. We examine various aspects of the proposed model in different experimental settings such as object localization, object segmentation, task priming, compact neural representation, and contextual interference reduction. We test the hypothesis that the proposed approach is capable of accomplishing hierarchical feature localization according to task cuing. Additionally, feature modulation using the proposed approach is tested for demanding tasks such as segmentation and iterative parameter fine-tuning. Moreover, the top-down attentional traces are harnessed to enable a more compact neural representation. The experimental achievements support the practical complementary role of the top-down selection mechanisms to the bottom-up feature encoding routines.