Text-to-Image diffusion models have shown remarkable capabilities in generating high-quality images. However, current models often struggle to adhere to the complete set of conditions specified in the input text and return unfaithful generations. Existing works address this problem by either fine-tuning the base model or modifying the latent representations during the inference stage with gradient-based updates. Not only are these approaches computationally expensive, but also they usually only improve limited kinds of errors (e.g., the count of objects). In this work, we propose an intervention-based mechanism to enhance the faithfulness of diffusion models by controlling the denoising process. Starting with layout-conditional diffusion models, our approach first detects incorrectly-generated/missing objects during denoising steps. Next, a layout is constructed from the erroneous objects (feedback). Finally, we return to an earlier denoising step. The new layout is fed to the diffusion model to obtain its latent representation. Correction is applied by composing the new latents with the original ones and continuing the generation process, thereby driving the generation away from erroneous directions. As additional feedback and correction strategy, we also explore retrieval-augmented generation to help the model recover missing objects. We conduct experiments on VPEval and HRS-Bench datasets and measure faithfulness across four dimensions; presence of objects, object counts, scale of objects and spatial relations between objects. Compared to GLIGEN, the state-of-the-art model on the VPEval dataset, our approach significantly improves on all metrics (+6.7% average accuracy increase). On HRS-Bench dataset, it also outperforms existing models in count and scale metrics.

Mingyu Ma, Jiun-Yu Kao, Arpit Gupta, Yu-Hsiang Lin, Wenbo Zhao, Tagyoung Chung, Wei Wang, Kai-Wei Chang, Nanyun Peng. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

As network technology evolves, telecom operators provide a variety of home network services such as Internet Protocol television (IPTV), voice over Internet Protocol (VoIP), video on demand (VoD), and other smart home solutions have become common. These services have led operators to expand the management scope from solely home gateway to a more diverse customer premises equipment (CPE), including set-top boxes (STBs) and home mesh Wi-Fi routers. These devices typically acquire an Internet Protocol (IP) address through IP over Ethernet (IPoE) and use the CPE WAN Management Protocol (CWMP) for remote management. However, operators cannot control customers’ improper operation of the devices, such as incorrect firewall configuration or wrong physical network connection. These issues cause the device to be unable to obtain an IP address and, thereby, cannot be managed via CWMP. This paper presents a multipath management mechanism based on CWMP, which manages CPE in the IPoE-based managed network and maintains the devices that cannot obtain IP through IPoE utilizing the Internet path. Experimental results show that the management efficiency of devices using this new mechanism reaches 93%, significantly higher than the 69% efficiency achieved with the original CWMP standard alone.

Most image dehazing deep learning models target synthetic datasets of hazy images, resulting in not considering features in natural hazy images. Leveraging on depth attention with adaptation, we propose a novel dehazing network called depthUNet, that is focused on natural images. Utilizing the correlation between depth information and haze distribution, our network enhances its generalization performance on natural images. Furthermore, our method improves the PSNR for the non-homogeneous realistic haze dataset NH-HAZE from 20.66 (DeHamer's result) to 20.74, using only 1.6% of the parameters. Similarly, for the outdoor scenes realistic haze dataset O-HAZE, our method enhances the PSNR from 24.36 (MSBDN's result) to 25.77, with just 6% of the parameters. On natural road images with haze from dataset RTTS, our method improved the vehicle detection rate by 10% in terms of the R-squared value. In summary, our method outperforms all state-of-the-art methods while utilizing the least number of parameters.

While recent progress has significantly boosted few-shot classification (FSC) performance, few-shot object detection (FSOD) remains challenging for modern learning systems. Existing FSOD systems follow FSC approaches, ignoring critical issues such as spatial variability and uncertain representations, and consequently result in low performance. Observing this, we propose a novelDual-Awareness Attention (DAnA)mechanism that enables networks to adaptively interpret the given support images. DAnA transforms support images intoquery-position-aware(QPA) features, guiding detection networks precisely by assigning customized support information to each local region of the query. In addition, the proposed DAnA component is flexible and adaptable to multiple existing object detection frameworks. By adopting DAnA, conventional object detection networks, Faster R-CNN and RetinaNet, which are not designed explicitly for few-shot learning, reach state-of-the-art performance in FSOD tasks. In comparison with previous methods, our model significantly increases the performance by 47% (+6.9 AP), showing remarkable ability under various evaluation settings.

In this letter, we predict the locations as probability distributions for the tasks of image object detection. We adopt the Kullback-Leibler divergences as the regression losses to train the deep neural networks. Since most existing evaluations label the objects with rectangular bounding boxes, we propose the Nearest Distribution Converter to find the closest uniform distributions from the predicted ones. Our proposed method can improve the detected accuracy measured in mAP by 0.57%, 0.75%, and 0.48% on the models YOLOv3, the YOLOv4-tiny, and the YOLOv4, respectively.

The equal-cost multi-path routing (ECMP) [4] achieves load balance in data centers network. Without network’s congestion status, ECMP may cause significant imbalance between paths. In this paper, we propose a better congestion aware routing protocol for Software Defined Network (SDN) to provide a better average link utilization. We follow the idea of In-band Network Telemetry (INT) to collect link congestion status in data center networks. Edge switches are responsible for detecting elephant flows by running a heavy hitter detection algorithm. When an elephant flow is reported to the controller by an edge switch, controller will use the collected congestion status to find the least congested path. In order to make the switches forward packets more efficiently and reduce the number of rules in switches’ forwarding table, we adopt label switching. We develop a Programming Protocol-independent Packet Processors (P4) program to design our novel routing scheme, which contains a heavy hitter detection algorithm. We further validate that our heavy hitter detection algorithm can run on Banzai machine. We also write a Python controller to communicate with P4 switches through P4 Runtime protocol. Our experimental results shows that the probing process in CAMP minimizes the bandwidth overhead in data centers. We use Mininet to construct fat-tree topologies and the emulated software P4switches run BMv2. The data mining workload is used to generate the traffic in our experiment. CAMP achieves better FCT compared to ECMP and HULA [6]. Also, the number of routing rules in CAMP maintains the smallest when network grows.

Quantitative differential phase-contrast (qDPC) imaging is a label-free phase retrieval method for weak phase objects using asymmetric illumination. However, qDPC imaging with fewer intensity measurements leads to anisotropic phase distribution in reconstructed images. In order to obtain isotropic phase transfer function, multiple measurements are required; thus, it is a time-consuming process. Here, we propose the feasibility of using deep learning (DL) method for isotropic qDPC microscopy from the least number of measurements. We utilize a commonly used convolutional neural network namely U-net architecture, trained to generate 12-axis isotropic reconstructed cell images (i.e. output) from 1-axis anisotropic cell images (i.e. input). To further extend the number of images for training, the U-net model is trained with a patch-wise approach. In this work, seven different types of living cell images were used for training, validation, and testing datasets. The results obtained from testing datasets show that our proposed DL-based method generates 1-axis qDPC images of similar accuracy to 12-axis measurements. The quantitative phase value in the region of interest is recovered from 66% up to 97%, compared to ground-truth values, providing solid evidence for improved phase uniformity, as well as retrieved missing spatial frequencies in 1-axis reconstructed images. In addition, results from our model are compared with paired and unpaired CycleGANs. Higher PSNR and SSIM values show the advantage of using the U-net model for isotropic qDPC microscopy. The proposed DL-based method may help in performing high-resolution quantitative studies for cell biology.

Yu-Hsiang Lin, Chian-Yu Chen, Jean Lee, Zirui Li, Yuyan Zhang, Mengzhou Xia, Shruti Rijhwani, Junxian He, Zhisong Zhang, Xuezhe Ma, Antonios Anastasopoulos, Patrick Littell, Graham Neubig. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019.

Small cells play an important role in the wireless networks for enhancing the radio coverage and capacity. However, to centrally monitor the quality of services (QoS) of several small cell access points (APs) from different equipment providers is very difficult. Therefore, a QoS monitoring mechanism based on TR-069 protocol is proposed to perform the performance management (PM) and configuration management (CM) of small cells in accordance with the data models in the reports of TR-196i2, TR-262 and TR-181. The key performance indicators (KPIs) of long term evolution (LTE) defined by 3GPP are used to evaluate the performance of each small cell and examine periodically statistical reports. According to the small cell's KPI, QoS monitoring mechanism sets up the dynamic threshold based on the weighted moving average. Furthermore, a transmission optimization mechanism based on the Elasticsearch distributed platform is proposed and implemented to analyze the 2,000 measurement collection (MC) files within 30 seconds for generating KPI reports of small cells. The telecom operator can refer to these reports for monitoring the real-time status of each small cell and improving the QoS of wireless networks.

Deep learning involves a difficult nonconvex optimization problem with a large number of weights between any two adjacent layers of a deep structure. To handle large data sets or complicated networks, distributed training is needed, but the calculation of function, gradient, and Hessian is expensive. In particular, the communication and the synchronization cost may become a bottleneck. In this letter, we focus on situations where the model is distributedly stored and propose a novel distributed Newton method for training deep neural networks. By variable and feature-wise data partitions and some careful designs, we are able to explicitly use the Jacobian matrix for matrix-vector products in the Newton method. Some techniques are incorporated to reduce the running time as well as memory consumption. First, to reduce the communication cost, we propose a diagonalization method such that an approximate Newton direction can be obtained without communication between machines. Second, we consider subsampled Gauss-Newton matrices for reducing the running time as well as the communication cost. Third, to reduce the synchronization cost, we terminate the process of finding an approximate Newton direction even though some nodes have not finished their tasks. Details of some implementation issues in distributed environments are thoroughly investigated. Experiments demonstrate that the proposed method is effective for the distributed training of deep neural networks. Compared with stochastic gradient methods, it is more robust and may give better test accuracy.

In order to keep reducing the bit cost, NAND Flash memory vendors have changed the NAND Flash technology from 2D to 3D since 2014. Moreover, 3D NAND Flash is becoming the mainstream of the NAND Flash based storage system from 2017. Owing to the storage material of NAND Flash changing from heavy doped poly-silicon to silicon-nitride, the inter-cell interference is ignored during programming operation to increase the write performance on 3D NAND Flash memory. However, in order to reduce the aspect ratio of the memory hole, the thickness of the inter word-line dielectric is reduced with the increasing of the stacking number. Accordingly, the Vth distribution is widened by the interference between the electrons in the silicon-nitride and the conductive channel between memory cells. This paper provides the measurement results of the cell-to-cell interference with the mass-produced 3D NAND Flash memory and proposes a method to reduce the cell-to-cell interference on 3D NAND Flash. Furthermore, the error bit of the 3D NAND Flash memory is 15% reduced and the endurance is 3X increased by the proposed method.

Wireless sensor networks (WSNs) are one of the most important ingredients in the Internet of Things. Thus it is vital to design a good power saving protocol, which operates at the medium access control (MAC) layer, for a WSN since sensors are generally battery-powered. On the other hand, organizing a WSN into coronas centered at the sink is a simple effective technique to achieve low-overhead routing where every sensor needs neither to broadcast beacons nor to maintain routing/neighbor tables. Hence in this paper, we propose an optimal and maximized configurable power saving protocol, named Green-MAC, for a corona-based WSN, which has the following attractive features. (i) By using the generalized Chinese remainder theorem, Green-MAC guarantees that any two sensors in the neighboring coronas can simultaneously wake up in bounded time regardless of their schedule offset as well as their respective cycle lengths. (ii) Given the cycle length, the ATF-ratio (i.e. the fraction of awake time frames in a cycle) of each sensor reaches the theoretical minimum. (iii) Under the minimum ATF-ratio constraints, the number of configurable ATF-ratios of each sensor reaches the theoretical maximum. (iv) An ATF-ratio configuration scheme is proposed for Green-MAC such that the power consumption of a WSN can be minimized while the worst event-to-sink delay requirement can be fulfilled with high probability. Both theoretical analysis and simulation results demonstrate that Green-MAC greatly outperforms existing power saving protocols for corona-based WSNs, including Q-MAC and Queen-MAC, in terms of ATF-ratio, configurability, network lifetime, delay violation ratio, and event-to-sink throughput.

Leakage tests have been a challenge for through-silicon vias (TSVs) in a 3-D IC. Most existing methods are still inadequate in terms of the range of testable leakage currents. In this paper, we borrow the wisdom of the IO-pin leakage test while enhancing it with two features. First, we make it more suitable for a TSV, which has a much smaller capacitance than an IO pin. Second, we support a wide range of leakage test (e.g., from 0.125 μA to 16 μA), and thereby allowing for flexible test threshold setting and leakage characterization. To achieve this goal, we present two sets of techniques-1) wait-time generation by programmable delay line, and 2) wait-time propagation with a self-timed timing control scheme to overcome the timing skew problem due to signal routing. We demonstrate that the entire scheme can be done in only logic gates, making it easy to integrate into the common design flow.

A parametric delay fault could arise in a through-silicon via (TSV) of a 3-D IC due to a manufacturing defect. Identification of such a fault is essential for fault diagnosis, yield-learning, and/or reliability screening. In this paper, we present an innovative design-for-testability technique called variable output thresholding. We discovered that by dynamically switching the output of a TSV from a normal inverter to a Schmitt-Trigger inverter, the parametric delay fault on the TSV can be characterized and detected. SPICE simulation reveals that this technique remains effective even when there is significant process variation. A scalable test infrastructure indicates that the test time is modest at only 17.2 ms for 1024 TSVs and 648.8 ms for 32768 TSVs when the test clock is running at 10 MHz.

In this paper, we propose a method and the required architecture for characterizing the propagation delays of the through Silicon vias (TSVs) in a 3-D IC. First of all, every two TSVs are paired up to form an oscillation ring with some peripheral circuits. Their joint performance can thus be measured roughly by the oscillation period of the ring. Next, we utilize a technique called sensitivity analysis to further derive the propagation delay of each individual TSV participating in an oscillation ring-a distilling process. In this process, we perturb the strength of the two TSV drivers, and then measure their effects in terms of the change of the oscillation ring's period. By some following analysis, the propagation delay of each TSV can be revealed. On top of scheme, we also present an architecture that can activate the performance characterization process of each test unit - that consists of two TSVs - one at a time in a proper sequence. The area overhead is only 18.97 equivalent two-input NAND gate per TSV, by which one can gain the ability to profile the capacitances and the propagation delays of the TSVs on a 3-D IC.

Leakage test has been a challenge for TSVs in a 3D IC. Most existing methods are still inadequate in terms of the range of leakage currents they can test. In this work, we borrow the wisdom of the IO pin leakage test while enhancing it with two features: (1) we make it more suitable for TSVs which has a much smaller capacitance than an IO pin, and (2) we support leakage binning in a wide range of currents from 1 uA to 128 uA, and thereby allowing flexible test threshold settings and leakage characterization. Since we use only logic gates in the Design-for-Testability circuit, it is also easier to be integrated into the TSV design flow than previous methods.

In this work, we present a robust small delay test scheme for through-silicon vias (TSVs) in a 3D IC. By changing the output inverter's threshold of a TSV in a testable oscillation ring structure, we can approximate the propagation delay across that TSV, and thereby detecting a small delay fault. SPICE simulation reveals that this Variable Output Thresholding (VOT) technique is still effective even when there is significant process variation in detecting a slow TSV with some resistive open defect that may escape the traditional at-speed test.

A TSV in a 3D IC could suffer from two major types of parametric faults — a resistive open fault, or a leakage fault. Dealing with these parametric faults (which do not destroy the functionality of a TSV completely but only degrade its quality or performance) is often trickier than dealing with a stuck-at fault. Previous works have not proposed a unified test structure and method that can characterize their respective effects. Based on our previous test structure, called VOT (Variable Output Threshold) scheme for delay faults, we propose a unified in-situ characterization flow for both parametric fault types of a post-bond TSV. With this flow, one can easily derive a more insightful assessment of a parametric fault in production test, process monitoring, and/or diagnosis-driven yield learning.

IEEE 802.11 is currently the most popular medium access control (MAC) standard for mobile ad hoc networks (MANETs). On the other hand, clustering in MANETs is a promising technique to ensure the scalability of various communication protocols. Thus, we propose an optimal asymmetric and maximized adaptive power management protocol, called OAMA, for 802.11-based clustered MANETs, which has the following attractive features. 1) Given the length of schedule repetition interval (SRI), the duty cycles of both clusterheads and members reach the theoretical minimum. 2) Under the minimum duty cycle constraints, the numbers of tunable SRIs for clusterheads and members reach the theoretical maximum. 3) By means of factor-correlative coterie-plane product, OAMA guarantees bounded-time neighbor discovery between the clusterhead and its member, and between all clusterheads, regardless of stations' individual SRIs and the schedule offset between neighboring stations. 4) The time complexity of OAMA neighbor maintenance is O(1). 5) OAMA adopts a cross-layer SRI adjustment scheme such that stations can adaptively tune the values of SRI to maximize energy conservation according to flow timeliness requirements. Both theoretical analyses and simulation results show that OAMA substantially outperforms existing power management protocols for clustered MANETs, including AQEC [2] and ACQ [14], in terms of duty cycle, adaptiveness, data delay dropped ratio, network lifetime, and end-to-end energy throughput.

IEEE 802.16 was created to meet the need of highspeed wireless access in metropolitan-scale areas. Due to air interference and dynamic queue state changes in subscriber stations, the idling UL-subframe problem, uplink hole problem, and padding waste problem are inevitable in 802.16 point-to- multipoint networks. To the best of our knowledge, this is the first work that seriously studies these issues. In this paper, we proposed the UBAR protocol, which employs the proportionally fair sharing scheme to utilize bandwidth efficiently, and adopts the timeout-based UL-MAP retransmission scheme with uplink bandwidth reallocation algorithms to simultaneously solve three bandwidth waste problems. Simulation results reveal that the uplink good put of UBAR in error-prone environments can be very close to that in an error-free environment.