No abstract

Classifiers such as deep neural networks have been shown to be vulnerable against adversarial perturbations on problems with high-dimensional input space. While adversarial training improves the robustness of image classifiers against such adversarial perturbations, it leaves them sensitive to perturbations on a non-negligible fraction of the inputs. In this work, we show that adversarial training is more effective in preventing universal perturbations, where the same perturbation needs to fool a classifier on many inputs. Moreover, we investigate the trade-off between robustness against universal perturbations and performance on unperturbed data and propose an extension of adversarial training that handles this trade-off more gracefully. We present results for image classification and semantic segmentation to showcase that universal perturbations that fool a model hardened with adversarial training become clearly perceptible and show patterns of the target scene.

This article focuses on the use of data gloves for human-computer interaction concepts, where external sensors cannot always fully observe the user's hand. A good concept hereby allows to intuitively switch the interaction context on demand by using different hand gestures. The recognition of various, possibly complex hand gestures, however, introduces unintentional overhead to the system. Consequently, we present a data glove prototype comprising a glove-embedded gesture classifier utilizing data from Inertial Measurement Units (IMUs) in the fingertips. In an extensive set of experiments with 57 participants, our system was tested with 22 hand gestures, all taken from the French Sign Language (LSF) alphabet. Results show that our system is capable of detecting the LSF alphabet with a mean accuracy score of 92% and an F1 score of 91%, using complementary filter with a gyroscope-to-accelerometer ratio of 93%. Our approach has also been compared to the local fusion algorithm on an IMU motion sensor, showing faster settling times and less delays after gesture changes. Real-time performance of the recognition is shown to occur within 63 milliseconds, allowing fluent use of the gestures via Bluetooth-connected systems.

No abstract

Deep neural networks achieve state-of-the-art results on several tasks while increasing in complexity. It has been shown that neural networks can be pruned during training by imposing sparsity inducing regularizers. In this paper, we investigate two techniques for group-wise pruning during training in order to improve network efficiency. We propose a gating factor after every convolutional layer to induce channel level sparsity, encouraging insignificant channels to become exactly zero. Further, we introduce and analyse a bounded variant of the 1 regularizer, which interpolates between 1 and 0-norms to retain performance of the network at higher pruning rates. To underline effectiveness of the proposed methods, we show that the number of parameters of ResNet-164, DenseNet-40 and MobileNetV2 can be reduced down by 30%, 69%, and 75% on CIFAR100 respectively without a significant drop in accuracy. We achieve state-of-the-art pruning results for ResNet-50 with higher accuracy on ImageNet. Furthermore, we show that the light weight Mo-bileNetV2 can further be compressed on ImageNet without a significant drop in performance. IntroductionModern deep neural networks are notoriously known for requiring large computational resources, which becomes particularly problematic in resource-constrained domains, such as in automotive, mobile or embedded applications. Neural network compression methods aim at reducing the computational footprint of a neural network while preserving task performance (e.g. classification accuracy) [4,34]. One family of such methods, Network pruning, operates by removing unnecessary weights or even whole neurons or convolutional featuremaps ("channels") during or after training, thus reducing computational resources needed at test time or deployment. A simple relevance-criterion for pruning weights is weight-magnitude: "small" weights contribute relatively little to the overall computation (dot-products and convolutions) and can thus be removed.However, weight-pruning leads to unstructured sparsity in weight matrices and filters. While alleviating storage demands, it is non-trivial to exploit