This paper introduces OpenMax, a novel method for open set recognition using deep networks. Deep networks are traditionally designed for closed set recognition, where all classes are known in advance. However, real-world recognition requires open set recognition, where the system must reject unknown or unseen classes during testing. OpenMax extends the SoftMax layer of deep networks to estimate the probability that an input belongs to an unknown class, enabling formal open set recognition. The key idea of OpenMax is to adapt Meta-Recognition concepts to the activation patterns in the penultimate layer of the network. By analyzing the activation vectors (AVs) of the network, OpenMax estimates the probability that an input is from an unknown class. This is done by computing the distance between the AV of an input and the model AVs of known classes. If the distance is large, the input is likely to be from an unknown class and is rejected. OpenMax is shown to significantly outperform basic deep networks and deep networks with thresholding of SoftMax probabilities in open set recognition. The method is evaluated using pre-trained networks from the Caffe Model Zoo on ImageNet 2012 validation data, as well as thousands of fooling and open set images. The results show that OpenMax effectively rejects fooling and open set images while maintaining high accuracy on known classes. The paper also discusses the challenges of open set recognition, including the difficulty of distinguishing between fooling images and open set images. Adversarial images, which are visually similar to training samples but are misclassified by deep networks, are particularly challenging. OpenMax is shown to reject such images by measuring the distance between the AV of an input and the model AVs of known classes. The paper presents an algorithm for OpenMax, which includes steps for computing mean activation vectors (MAVs) for each class, fitting a Weibull distribution to the distances between the AVs of correctly classified training examples and the MAVs, and using these parameters to estimate the probability that an input is from an unknown class. The algorithm is evaluated on the ILSVRC 2012 dataset, showing that OpenMax achieves higher F-measure performance than SoftMax in open set recognition tasks. The paper also discusses the limitations of OpenMax, including the need for careful calibration of parameters and the potential for increased false rejection of true classes when using larger tail sizes in EVT calibration. However, the results show that OpenMax provides a formal solution to open set recognition by bounding the open space risk and allowing the system to reject unknown classes.This paper introduces OpenMax, a novel method for open set recognition using deep networks. Deep networks are traditionally designed for closed set recognition, where all classes are known in advance. However, real-world recognition requires open set recognition, where the system must reject unknown or unseen classes during testing. OpenMax extends the SoftMax layer of deep networks to estimate the probability that an input belongs to an unknown class, enabling formal open set recognition. The key idea of OpenMax is to adapt Meta-Recognition concepts to the activation patterns in the penultimate layer of the network. By analyzing the activation vectors (AVs) of the network, OpenMax estimates the probability that an input is from an unknown class. This is done by computing the distance between the AV of an input and the model AVs of known classes. If the distance is large, the input is likely to be from an unknown class and is rejected. OpenMax is shown to significantly outperform basic deep networks and deep networks with thresholding of SoftMax probabilities in open set recognition. The method is evaluated using pre-trained networks from the Caffe Model Zoo on ImageNet 2012 validation data, as well as thousands of fooling and open set images. The results show that OpenMax effectively rejects fooling and open set images while maintaining high accuracy on known classes. The paper also discusses the challenges of open set recognition, including the difficulty of distinguishing between fooling images and open set images. Adversarial images, which are visually similar to training samples but are misclassified by deep networks, are particularly challenging. OpenMax is shown to reject such images by measuring the distance between the AV of an input and the model AVs of known classes. The paper presents an algorithm for OpenMax, which includes steps for computing mean activation vectors (MAVs) for each class, fitting a Weibull distribution to the distances between the AVs of correctly classified training examples and the MAVs, and using these parameters to estimate the probability that an input is from an unknown class. The algorithm is evaluated on the ILSVRC 2012 dataset, showing that OpenMax achieves higher F-measure performance than SoftMax in open set recognition tasks. The paper also discusses the limitations of OpenMax, including the need for careful calibration of parameters and the potential for increased false rejection of true classes when using larger tail sizes in EVT calibration. However, the results show that OpenMax provides a formal solution to open set recognition by bounding the open space risk and allowing the system to reject unknown classes.