Toward Latent Cognizance on Open-Set Recognition

Open-Set Recognition (OSR) has been actively studied recently. It attempts to address a closed-set paradigm of conventional object recognition. Most OSR approaches are quite analytic and retrospective, associable to human's system-2 decision. A novel bayesian-based approach Latent Cognizance (LC), derived from a new probabilistic interpretation of softmax output, is more similar to natural impulse response and more associable to system-1 decision. As both decision systems are crucial for human survival, both OSR approaches may play their roles in development of machine intelligence. Although the new softmax interpretation is theoretically sound and has been experimentally verified, many progressive assumptions underlying LC have not been directly examined. Our study clarifies those assumptions and directly examines them. The assumptions are laid out and tested in a refining manner. The investigation employs AlexNet and VGG as well as ImageNet and Cifar-100 datasets. Our findings support the existence of the common cognizance function, but the evidence is against generality of a common cognizance function across base models or application domains.