Search

The SAS-CBRS framework is being tested to share the federally held spectrum with licensed users and opportunistic users to maximize the underutilized spectrum’s utility and overcome spectrum scarcity. In the SAS-CBRS framework, radio resources are assigned to the incumbent access (IA), primary access licensees (PAL), and general authorized access (GAA) users according to the given priority. The SAS-CBRS three-tier framework is different from the conventional cognitive radio networks (CRN) as it involves a central entity that acts as a server called a spectrum access system (SAS). The methods to assign the resources using the SAS are still in the research phase. Yet, no standard method is defined by the FCC for resource allocation. The current CRN methods cannot be directly applied because of the addition of the third tier and a central server. Moreover, strict rules are defined for using the 3.5 GHz spectrum band for communication. In this paper, a novel DDRQ-SAS algorithm integrated with the double auction (DA) algorithm is proposed that uses deep recurrent double Q-learning. The DDRQ-SAS is used by the SAS to hold a spectrum auction and create a spectrum pool to get information on PAL channels. PAL operators use the DA algorithm to generate the asking prices intelligently for their available idle channels and the GAA users will use the DA algorithm to intelligently bid for their preferred channels. The DDRQ-SAS-DA algorithm allows the GAA users to get the guaranteed QoS offered by the PAL operators in an auction. GAA users maintain the preference list of the PAL reserved idle channels and bid intelligently based on the available QoS. SAS completes the transaction by allocating the channels to the winning GAAs. The defined problem is also modeled using the double auction multi-winner multi-channel technique and the TDSA-PS algorithm. Numerical results show that the proposed DDRQ-SAS-DA algorithm provides up to 20% better QoS at higher loads for GAA users, generates 24% more revenue for PAL operators, and is 1.6 times more efficient in assigning 500 GAA users.

Accurate detection and classification of wafer defects constitute an important component in semiconductor manufacturing. It provides interpretable information to find the possible root causes of defects and to take actions for quality management and yield improvement. Traditional approach to classify wafer defects, performed manually by experienced engineers using computer-aided tools, is time-consuming and can be low in accuracy. Hence, automated detection of wafer defects using deep learning approaches has attracted considerable attention to improve the performance of detection process. However, a majority of these works have focused on defect classification and have ignored defect localization which is equally important in determining how specific process steps can lead to defects in certain locations. To address this, we evaluate the state-of-the-art You Only Look Once (YOLO) architecture to accurately locate and classify wafer map defects. Experimental results obtained on 19200 wafer maps show that YOLOv3 and YOLOv4, the variants of YOLO architecture, can achieve >94% of classification accuracy in real-time. For comparison, other architectures, namely ResNet50 and DenseNet121 are also evaluated for wafer defect classification and they give accuracies 89% and 92% respectively, however, without localization abilities. We find that the object detection methods are very useful in locating and classifying defects on semiconductor wafers.

Deep learning models hold state of the art performance in many fields, yet their design is still based on heuristics or grid search methods that often result in overparametrized networks. This work proposes a method to analyze a trained network and deduce an optimized, compressed architecture that preserves accuracy while keeping computational costs tractable. Model compression is an active field of research that targets the problem of realizing deep learning models in hardware. However, most pruning methodologies tend to be experimental, requiring large compute and time intensive iterations of retraining the entire network. We introduce structure into model design by proposing a single shot analysis of a trained network that serves as a first order, low effort approach to dimensionality reduction, by using PCA (Principal Component Analysis). The proposed method simultaneously analyzes the activations of each layer and considers the dimensionality of the space described by the filters generating these activations. It optimizes the architecture in terms of number of layers, and number of filters per layer without any iterative retraining procedures, making it a viable, low effort technique to design efficient networks. We demonstrate the proposed methodology on AlexNet and VGG style networks on the CIFAR-10, CIFAR-100 and ImageNet datasets, and successfully achieve an optimized architecture with a reduction of up to 3.8X and 9X in the number of operations and parameters respectively, while trading off less than 1% accuracy. We also apply the method to MobileNet, and achieve 1.7X and 3.9X reduction in the number of operations and parameters respectively, while improving accuracy by almost one percentage point.

Recent advances in deep learning, coupled with the onslaught of unlabelled medical data have drawn ever-increasing research interests by discovering multiple levels of distributed representations and solving complex medical related problems. Malaria disease detection in early stage requires an accurate and precise diagnosis in order to achieve successful patient remission. This paper proposes a comprehensive computer-aided diagnosis (CAD) scheme for identifying the presence of malaria parasites in thick blood smear images. The parameters of the scheme are pre-trained by functional link artificial neural network followed by sparse stacked autoencoder. The optimum size of the CAD scheme used in this research is 12500-2500-100-50-2, where the input layer has 12500 nodes and Softmax classifier output layer has 2 nodes. Moreover, the 10- fold cross validation reflects that the classification is reliable and is applicable to new patient blood smear images. The proposed CAD scheme has been evaluated using malaria blood smear image data set, achieving a detection accuracy of 89.10%, a sensitivity of 93.90% and specificity of 83.10%. The extensive comparative experiment suggests that the proposed CAD scheme provides richer effectiveness and efficiency for malaria data set compared to other deep learning techniques for better diagnosis decision and management. This work implements a novel approach to fast processing and will be a beneficial tool in disease identification.

Adversarial examples are perturbed inputs that are designed (from a deep learning network's (DLN) parameter gradients) to mislead the DLN during test time. Intuitively, constraining the dimensionality of inputs or parameters of a network reduces the “space”in which adversarial examples exist. Guided by this intuition, we demonstrate that discretization greatly improves the robustness of the DLNs against adversarial attacks. Specifically, discretizing the input space (or allowed pixel levels from 256 values or 8bit to 4 values or 2bit) extensively improves the adversarial robustness of the DLNs for a substantial range of perturbations for minimal loss in test accuracy. Furthermore, we find that binary neural networks (BNNs) and related variants are intrinsically more robust than their full precision counterparts in adversarial scenarios. Combining input discretization with the BNNs furthers the robustness, even waiving the need for adversarial training for the certain magnitude of perturbation values. We evaluate the effect of discretization on MNIST, CIFAR10, CIFAR100, and ImageNet datasets. Across all datasets, we observe maximal adversarial resistance with 2bit input discretization that incurs an adversarial accuracy loss of just ~ 1% - 2% as compared to clean test accuracy against single-step attacks. We also show standalone discretization remains vulnerable to stronger multi-step attack scenarios necessitating the use of adversarial training with discretization as an improved defense strategy.