Towards Open Set Deep Networks (1511.06233v1)

Abstract: Deep networks have produced significant gains for various visual recognition problems, leading to high impact academic and commercial applications. Recent work in deep networks highlighted that it is easy to generate images that humans would never classify as a particular object class, yet networks classify such images high confidence as that given class - deep network are easily fooled with images humans do not consider meaningful. The closed set nature of deep networks forces them to choose from one of the known classes leading to such artifacts. Recognition in the real world is open set, i.e. the recognition system should reject unknown/unseen classes at test time. We present a methodology to adapt deep networks for open set recognition, by introducing a new model layer, OpenMax, which estimates the probability of an input being from an unknown class. A key element of estimating the unknown probability is adapting Meta-Recognition concepts to the activation patterns in the penultimate layer of the network. OpenMax allows rejection of "fooling" and unrelated open set images presented to the system; OpenMax greatly reduces the number of obvious errors made by a deep network. We prove that the OpenMax concept provides bounded open space risk, thereby formally providing an open set recognition solution. We evaluate the resulting open set deep networks using pre-trained networks from the Caffe Model-zoo on ImageNet 2012 validation data, and thousands of fooling and open set images. The proposed OpenMax model significantly outperforms open set recognition accuracy of basic deep networks as well as deep networks with thresholding of SoftMax probabilities.

• The paper presents OpenMax to overcome deep networks’ limitations in open-set recognition by incorporating Weibull-based meta-recognition.
• It computes mean activation vectors per class and adjusts activation scores to effectively reject unknown and misleading inputs.
• Empirical results on ImageNet, combined with theoretical proofs, confirm OpenMax’s capability for robust, real-world visual recognition.

Towards Open Set Deep Networks

The research paper titled "Towards Open Set Deep Networks" by Abhijit Bendale and Terrance E. Boult addresses a crucial limitation in the deployment of deep networks for real-world visual recognition tasks. While deep networks have achieved substantial success in closed-set recognition scenarios, they inherently face challenges when applied to open-set environments where instances from unseen classes may be encountered. This limitation has practical ramifications, as real-world applications frequently encounter data from classes not represented in the training set.

Key Contributions

The paper proposes a novel methodology to adapt deep networks for open-set recognition. The core contributions are:

1. Introduction of OpenMax Layer: The OpenMax layer is a replacement for the traditional SoftMax layer. This new layer estimates the probability of an input belonging to an unknown class by leveraging meta-recognition concepts.
2. Meta-Recognition Using Activation Vectors: The methodology utilizes activation patterns in the penultimate layer of the network to estimate the likelihood of an input example being from an unknown class. This is achieved by fitting Weibull distributions to the distances between activation vectors and the mean activation vectors of known classes.
3. Theoretical Framework: The authors provide a theoretical proof demonstrating that their OpenMax approach manages open space risk, thereby formally supporting open set recognition.

Methodology

The OpenMax approach involves several steps:

1. Activation Vector Computation: During the training phase, mean activation vectors (MAV) are computed for each class using correctly classified examples. The distances from these MAVs are then modeled using Weibull distributions.
2. OpenMax Calculation: At test time, for any input image, the activation vector is computed and distances to the MAVs are measured. The Weibull parameters are used to compute weights that adjust the activation scores to include the probability of the input being from an unknown class.
3. Probability Estimation: By including an unknown class and adjusting the activation scores, OpenMax provides probability estimates that can reject unknown inputs. The classification decision is based on these revised probabilities.

Experimental Results

The proposed OpenMax model was evaluated using the ImageNet 2012 dataset, including validation images from known classes, open-set images from unseen categories, and fooling images. The results highlight several significant findings:

• Improved Open Set Recognition: The OpenMax model significantly outperforms traditional deep networks with SoftMax layers in terms of open set recognition. It reduces the number of erroneous classifications of unknown and fooling inputs.
• Rejection of Fooling Images: OpenMax effectively mitigates the problem of fooling images, which are crafted to deceive the network into making high-confidence but incorrect classifications.
• Theoretical Validation: The paper's theoretical foundation supports the empirical findings, showing that OpenMax manages open space risk in feature space, thus qualifying as an open set recognition function.

Implications and Future Directions

The introduction of OpenMax has both practical and theoretical implications. Practically, it allows deep networks to be more robust and applicable in dynamic environments where unknown classes can frequently appear. This could enhance the deployment of deep learning models in security systems, autonomous driving, medical diagnostics, and other fields requiring high reliability under uncertain conditions.

Theoretically, the methods introduced open avenues for further enhancement of open set recognition. For example, exploring more sophisticated models for activation vectors, such as multiple exemplar-based models or leveraging metric learning, might further improve robustness and accuracy. Additionally, integrating OpenMax with adversarial training strategies could yield insights into mitigating adversarial attacks more effectively.

Conclusion

The paper "Towards Open Set Deep Networks" offers a substantial advancement in the adaptation of deep networks for open-set scenarios through its novel OpenMax layer and meta-recognition techniques. This research not only provides a practical solution to a significant real-world problem but also lays the groundwork for future explorations in robust and adaptive deep learning methodologies.