Tianyang Zhou, Ziyi Zhang, Haowen Lin, Somesh Jha, Mihai Christodorescu, Kirill Levchenko, Varun Chandrasekaran. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Recent advancements have marked significant progress in photoacoustic imaging as an effective method for acquiring deep bio-tissue visuals in modern medical clinical therapy and the efficacy of U-Net and its variants has been established for imaging quality enhancement in this field. Unlike common computer vision datasets such as ImageNet [1] and PASCAL VOC [2], biomedical images exhibit highly structured patterns, low spatial resolution, and single-channel modality, as shown in Fig. 1. Additionally, the U-Net parameters trained for medical super-resolution tasks demonstrate a high sparsity ratio, making them suitable for implementation on edge-computing platforms. Therefore, developing an energy-efficient photoacoustic imaging setup in this area is a natural progression. However, this development is constrained by the current neural network architectures, which are built around a U-Net backbone. The multi-stage feature extractor, skip connection integration across different blocks, and the encoder-decoder backbone design pose significant challenges to cutting-edge computational hardware platforms. In this study, a scalable, sparsity-supported neural network accelerator architecture for bio-tissue imaging quality enhancement is proposed to meet the stringent requirements of latency and energy efficiency, as depicted in Fig. 2. This architecture achieves desired performance improvements by exploring the sparsity possibilities in neural network during the training process and implementing an end-to-end pixel-first hardware design to minimize data movement and support sparsity computation. Compared with the state-of-the-art related works, this optimized architecture has achieved minimum on-chip storage overhead and the fastest frame for the application of photoacoustic imaging quality enhancement. The scalable architecture has also been implemented on a Xilinx XCZU9EG FPGA and attains a performance of PSNR@ 24 dB and a frame rate of 164 fps at a working frequency of 250 MHz.

Centralized training for Human Activity Recognition (HAR) typically relies heavily on vast amounts of aggregated data, compromising user privacy. Federated learning (FL) for HAR offers a solution to protect local data privacy. However, existing FL methodologies often fail to fully capture the heterogeneity of user data and the latent correlations among user models, resulting in suboptimal performance and limited robustness. This paper proposes a Correlation-Aware Multi-Similarty Learning Method for Federated HAR, namely MultiSim. Our approach enhances model accuracy with an effective inter-user knowledge learning while protecting data privacy. MultiSim first constructs multiple similarity metrics, and then makes model feature fusion cunningly by the above metrics to learn inherent user similarity profiles. Additionally, we introduce a novel clustering-based FL framework by isolating malicious nodes, thereby mitigating the impact of adversarial attacks. Extensive evaluations on two realworld HAR datasets demonstrate the superiority of MultiSim over other state-of-the-art FL methods under accuracy and robustness. These findings demonstrate MultiSim's potential as a robust and effective solution for HAR.

General Matrix Multiplication (GEMM) has a wide range of applications in scientific simulation and artificial intelligence. Although traditional libraries can achieve high performance on large regular-shaped G EMMs, they often behave not well on irregular-shaped G EMMs, which are often found in new algorithms and applications of high-performance computing (HPC). Due to energy efficiency constraints, low-power multi-core digital signal processors (DSPs) have become an alternative architecture in HPC systems. Targeting multi-core DSPs in FT-m7032, a prototype CPU-DSPs heterogeneous processor for HPC, an efficient implementation-ftIMM - for three types of irregular-shaped GEMMs is proposed. FtIMM supports automatic generation of assembly micro-kernels, two parallelization strategies, and auto-tuning of block sizes and parallelization strategies. The experiments show that ftIMM can get better performance than the traditional GEMM implementations on multi-core DSPs in FT-m7032, yielding on up to 7.2x performance improvement, when performing on irregular-shaped GEMMs. And ftIMM on multi-core DSPs can also far outperform the open source library on multi-core CPUs in FT-m7032, delivering up to 3.1 x higher efficiency.

By localizing microbubbles (MBs) in the vasculature, ultrasound localization microscopy (ULM) has recently been proposed, which greatly improves the spatial resolution of ultrasound (US) imaging and will be helpful for clinical diagnosis. Nevertheless, several challenges remain in fast ULM imaging. The main problems are that current localization methods used to implement fast ULM imaging, e.g., a previously reported localization method based on sparse recovery (CS-ULM), suffer from long data-processing time and exhaustive parameter tuning (optimization). To address these problems, in this paper, we propose a ULM method based on deep learning, which is achieved by using a modified sub-pixel convolutional neural network (CNN), termed as mSPCN-ULM. Simulations and in vivo experiments are performed to evaluate the performance of mSPCN-ULM. Simulation results show that even if under high-density condition (6.4 MBs/mm2), a high localization precision (~28 μm in the lateral direction and ~24 μm in the axial direction) and a high localization reliability (Jaccard index of 0.66) can be obtained by mSPCN-ULM, compared to CS-ULM. The in vivo experimental results indicate that with plane wave scan at a transmit center frequency of 15.625 MHz, microvessels with diameters of ~17 μm can be detected and adjacent microvessels with a distance of ~42 μm can be separated. Furthermore, when using GPU acceleration, the data-processing time of mSPCN-ULM can be shortened to ~6 sec/frame in the simulations and ~23 sec/frame in the in vivo experiments, which is 3-4 orders of magnitude faster than CS-ULM. Finally, once the network is trained, mSPCN-ULM does not need parameter tuning to implement ULM. As a result, mSPCN-ULM opens the door to implement ULM with fast data-processing speed, high imaging accuracy, short data-acquisition time, and high flexibility (robustness to parameters) characteristics.

Penetration testing offers strong advantages in the discovery of hidden vulnerabilities in a network and assessing network security. However, it can be carried out by only security analysts, which costs considerable time and money. The natural way to deal with the above problem is automated penetration testing, the essential part of which is automated attack planning. Although previous studies have explored various ways to discover attack paths, all of them require perfect network information beforehand, which is contradictory to realistic penetration testing scenarios. To vividly mimic intruders to find all possible attack paths hidden in a network from the perspective of hackers, we propose a network information gain based automated attack planning (NIG-AP) algorithm to achieve autonomous attack path discovery. The algorithm formalizes penetration testing as a Markov decision process and uses network information to obtain the reward, which guides an agent to choose the best response actions to discover hidden attack paths from the intruder’s perspective. Experimental results reveal that the proposed algorithm demonstrates substantial improvement in training time and effectiveness when mining attack paths.