The paper "Towards Open Set Deep Networks" by Abhijit Bendale and Terrance E. Boult addresses the limitations of deep networks in handling open set recognition, where the system must reject unknown or unseen classes. The authors propose a new model layer called OpenMax, which estimates the probability of an input being from an unknown class. OpenMax incorporates Meta-Recognition concepts to estimate the unknown probability by adapting activation patterns in the penultimate layer of the network. This approach allows the network to reject "fooling" and unrelated open set images, significantly reducing obvious errors. The paper demonstrates that OpenMax provides bounded open space risk and outperforms basic deep networks and those with thresholded SoftMax probabilities in open set recognition tasks. Experiments using pre-trained networks from the Caffe Model-zoo on ImageNet 2012 validation data show that OpenMax significantly improves accuracy in rejecting unknown classes, fooling images, and obvious errors from adversarial images while maintaining accuracy on testing images. The contributions of the paper include the development of Multi-class Meta-Recognition using Activation Vectors, the formalization of open set deep networks, and experimental validation of the effectiveness of OpenMax.The paper "Towards Open Set Deep Networks" by Abhijit Bendale and Terrance E. Boult addresses the limitations of deep networks in handling open set recognition, where the system must reject unknown or unseen classes. The authors propose a new model layer called OpenMax, which estimates the probability of an input being from an unknown class. OpenMax incorporates Meta-Recognition concepts to estimate the unknown probability by adapting activation patterns in the penultimate layer of the network. This approach allows the network to reject "fooling" and unrelated open set images, significantly reducing obvious errors. The paper demonstrates that OpenMax provides bounded open space risk and outperforms basic deep networks and those with thresholded SoftMax probabilities in open set recognition tasks. Experiments using pre-trained networks from the Caffe Model-zoo on ImageNet 2012 validation data show that OpenMax significantly improves accuracy in rejecting unknown classes, fooling images, and obvious errors from adversarial images while maintaining accuracy on testing images. The contributions of the paper include the development of Multi-class Meta-Recognition using Activation Vectors, the formalization of open set deep networks, and experimental validation of the effectiveness of OpenMax.