In image object detection, the training and inference of detectors based on random proposals are challenging due to their diverse, random and flexible nature. Existing work employs the use of generative diffusion models with subsequent processing of random proposals, which is computationally expensive and results in a slow detection speed. In order to address these issues, we propose a discrimination-based method for image object detection with random distinct proposals. A discriminative detection framework is employed, where in dense anchor frames are supplanted by random distinct proposals, with an emphasis on refining the training and inference strategies. In the training phase, the generated random proposals are fine-tuned to identify suitable offsets that disrupt the completely random distribution. In the inference phase, a novel evaluation strategy is employed to gradually refine a set of random distinct proposals into outputs. In comparison to previous state-of-the-art works, the metrics AP and others demonstrate significant improvement on the test benchmarks MS-COCO and CrowdHuman.

Blind Source Separation (BSS) separates mixed signals by leveraging source signals’ statistical independence and non-Gaussian characteristics, requiring no prior channel knowledge. As FastICA achieves rapid convergence, its complexity scales quadratically with signal dimensions. Although ICA-R enhances convergence via reference signals, iterative frameworks persist. We propose a low-complexity architecture: pre-analyzing mixing characteristics through correlation matrix decomposition, constructing reference-target correlation models, and deriving single-step closed-form solutions. The experiments show lower complexity compared to conventional methods, without merely performance loss.

Context. While models like BERT and GPT are powerful, they require substantial resources. Knowledge distillation can be employed as a technique to enhance their efficiency. Yet, we lack a clear understanding on their performance and energy consumption. This uncertainty is a major concern, especially in practical applications, where these models could strain resources and limit accessibility for developers with limited means. Our drive also comes from the pressing need for environmentally-friendly and sustainable applications in light of growing environmental worries. To address this, it is crucial to accurately measure their energy consumption. Goal. This study aims to determine how Knowledge Distillation affects the energy consumption and performance of NLP models. Method. We benchmark BERT, Distilled-BERT, GPT-2, and Distilled-GPT-2 using three different tasks from 3 different categories selected from a third-party dataset. The energy consumption, CPU utilization, memory utilization, and inference time of the considered NLP models are measured and statistically analyzed.

In recent years, the Transformer, leveraging its global attention mechanism, has demonstrated powerful performance in visual tasks such as image recognition and object detection. The application of visual Transformers for occlusion pedestrian detection has gained popularity as a promising research direction. Addressing shortcomings in existing detection algorithms and DETR series models for occlusion detection tasks, we propose HWA-DETR, an improved pedestrian detection algorithm. This enhancement incorporates height-width modulated attention and branching decoupling reparameterization from an application perspective. To boost detector performance, we directly process the image using the Transformer. The query inputs are designed with reparameterization of branch decoupling, significantly enhancing the model’s ability to recognize occluded pedestrian objects. During the re-coupling of the two branches, a high-width modulation attention mechanism is employed. This ensures that the model not only focuses on the content features of pedestrian objects but also accurately comprehends and predicts the spatial distributions of pedestrian objects in complex environments, particularly under occlusion conditions. These design improvements effectively enhance the model’s capability to handle occluded pedestrian detection with practical applications, achieving excellent detection performance on the occlusion subset of three datasets: Caltech, CityPersons, and EuroCityPersons. Additionally, extension experiments on two datasets, Caltech, CityPersons, demonstrate good robustness and generalization.

Multimodal emotion recognition has attracted increasing research attention due to its critical role in real applications. However, we frequently encounter the challenge of incomplete modalities, which adversely affects the accuracy of emotion recognition. Various methods have been proposed to address the issue of missing modalities, but few of them investigate multiple missing modalities with uncertain proportions and multimodal imbalance robustness existing in multimodal fusion. Moreover, existing approaches often fail to fully utilize high-level semantic information under both missing-modality and full-modality conditions. In this paper, we introduce a novel framework to tackle this issue, which consists of three key components: the LSTM modal feature encoding module, cascaded residual Encoder-Decoder reconstruction network based on CNN and Transformer, and gated attention fusion module. Extensive experiments on benchmark datasets IEMOCAP and MSP-IMPROV demonstrate that our model can effectively learn robust multimodal representations under different missing rates, compared with the comparison methods, the recognition performance is superior. In particular, our model achieves the highest weighted average F1-score (WAF) on IEMOCAP and the highest weighted average accuracy (WA) on MSP-IMPROV.

This paper presents a fast transient response capless low-dropout (LDO) regulator in 65 nm CMOS process for system-on-chip (SOC). The LDO regulator utilizes a feedforward ripple cancellation circuit (FFRCC) to achieve high power supply rejection (PSR), and a voltage damper is used to enhance transient response. Besides, a negative capacitance circuit (NCC) is added to the gate of the power stage to expand the bandwidth of FFRCC. The proposed LDO regulator is fabricated in 65 nm CMOS technology. Its voltage recovery time is 1.1 μs and 400 ns, respectively, when the load current steps from 200 μA to 50 mA or 50 mA to 200 μA with the rise/fall time of 100 ns. Its overshoot and undershoot voltages are 114 mV and 98 mV, respectively. Moreover, the regulator achieves -78 dB PSR at 2 MHz.

Ultrasound plane wave imaging is widely used in many applications thanks to its capability in reaching high frame rates. However, the amount of data acquisition and storage in a period of time can become a bottleneck in ultrasound system design for thousands frames per second. In our previous study, we proposed a low-rank and joint-sparse model to reduce the amount of sampled channel data of focused beam imaging by considering all the received data as a 2D matrix. However, for a single plane wave transmission, the number of channels is limited and the low-rank property of the received data matrix is no longer achieved. In this study, a L0-norm based Hankel structured low-rank and sparse model is proposed to reduce the channel data. An optimization algorithm, based on the alternating direction method of multipliers (ADMM), is proposed to efficiently solve the resulting optimization problem. The performance of the proposed approach was evaluated using the data published in Plane Wave Imaging Challenge in Medical Ultrasound (PICMUS) in 2016. Results on channel and plane wave data show that the proposed method is better adapted to the ultrasound channel signal and can recover the image with fewer samples than the conventional CS method.

The Internet’s congestion control landscape is currently in the midst of an unprecedented paradigm shift. A recent measurement study found that BBR, a congestion control algorithm introduced by Google in 2016, has seen rapid adoption and is deployed at more than 20% of the Alexa Top 20,000 websites. Encouraging early deployment results from Google, Dropbox and Spotify suggest that BBR could potentially replace traditional loss-based congestion control algorithms like CUBIC. In this paper, we study the interactions between CUBIC and BBR and show that the underlying interactions can be modeled as a normal form game. Our game-theoretic analysis and testbed measurements suggest that while BBR seems to achieve somewhat better performance than CUBIC on the Internet today, this advantage will decrease as the proportion of BBR flows increases. The distribution of congestion control algorithms on the Internet would likely reach a Nash Equilibrium, where no flow has the incentive to switch from CUBIC to BBR, or vice versa. We also found that the distribution of CUBIC and BBR flows in this Nash Equilibrium will be dependent mainly on the size of the bottleneck buffer, and marginally on the RTT distribution of the flows. Our results suggest that the future Internet will likely be more heterogeneous and that buffer sizing will continue to have a significant impact on Internet congestion control.

With the networks becoming complex, traditional congestion control protocols face increasing challenges in providing high-quality services for users. Traditional TCP and its variants fail to achieve high performance due to drawbacks in architectural design: predefined actions to specific network feedback. In this paper, we develop a learning-based TCP congestion control scheme RLCC, featuring a deep Q-network framework, in which senders learn the optimal control policies from observations instead of predefined rules. To apply DQN algorithms to congestion control problems, we first prove theoretically that congestion control problems are of Markov property. Therefore, the model-free reinforcement learning algorithm DQN can be used to solve congestion control. This is because the application of DQN to the network congestion control problem is convergent, and there exists an optimal strategy to obtain the best action for congestion control. We improved the network's performance by carefully designing the reward function and choosing the appropriate form and parameters through extensive experimentation. Extensive experiments on real-world environments of Pantheon via AWS confirm that RLCC can achieve utilization improvements over the traditional TCP congestion control schemes with higher throughput and lower transmission latency, and outperform the recently proposed learning-based congestion control protocol.

The large demand of mobile devices creates significant concerns about the quality of mobile applications (apps). Developers heavily rely on bug reports in issue tracking systems to reproduce failures (e.g., crashes). However, the process of crash reproduction is often manually done by developers, making the resolution of bugs inefficient, especially that bug reports are often written in natural language. To improve the productivity of developers in resolving bug reports, in this paper, we introduce a novel approach, called ReCDroid, that can automatically reproduce crashes from bug reports for Android apps. ReCDroid uses a combination of natural language processing (NLP) and dynamic GUI exploration to synthesize event sequences with the goal of reproducing the reported crash. We have evaluated ReCDroid on 51 original bug reports from 33 Android apps. The results show that ReCDroid successfully reproduced 33 crashes (63.5% success rate) directly from the textual description of bug reports. A user study involving 12 participants demonstrates that ReCDroid can improve the productivity of developers when resolving crash bug reports.

This paper presents a high-throughput and multi-parallel VLSI hardware architecture for the deblocking filter in the HEVC video coding standard. First, an implementation-friendly and fast boundary judgment method is proposed to avoid using the original recursion loop approach. Then a dedicated parallel VLSI architecture composed of four parallel filtering cores is presented based on the proposed boundary judgment method. With the parallel luma/chroma filtering and parallel vertical/horizontal edges filtering order, the proposed VLSI architecture can process filtering operations for one largest coding unit (LCU) with less filtering cycles than other conventional approaches. Furthermore, filtering efficiency is improved due to a novel ping-pang buffer architecture and the on-chip single-port SRAM with dedicated data arrangement in the memory modules. Experimental results demonstrate that the proposed deblocking filter architecture improves the performance by 28-89% at the expense of the slightly increased gate count compared to the previously known architecture in HEVC. The proposed architecture can reach a high operating clock frequency of 278 MHz with TSMC 90 nm library and meet the real time requirement of the deblocking filter for 8 K × 4 K video format at 123 frame/s.

This paper presents a novel VLSI hardware architecture for the real-time high-throughput implementation of the HEVC deblocking filtering. Based on the proposed implementation-friendly boundary judgment method, a dedicated multi-parallel architecture composed of four parallel filtering cores, parallel luma/chroma filtering and parallel vertical/horizontal edges filtering is presented. Experimental results demonstrate that the proposed architecture can greatly improve the performance at the expense of the slightly increased hardware cost compared to the previously known architecture in HEVC. The proposed architecture can also meet the real-time requirement of the deblocking filter for 8K×4K video format at 123fps under 278MHz clock rate.