Huaxiang Lu

FPGA accelerators for lightweight convolutional neural networks (LWCNNs) have recently attracted significant attention. Most existing LWCNN accelerators focus on single-Computing-Engine (CE) architecture with local optimization. However, these designs typically suffer from high on-chip/off-chip memory overhead and low computational efficiency due to their layer-by-layer dataflow and unified resource mapping mechanisms. To tackle these issues, a novel multi-CE-based accelerator with balanced dataflow is proposed to efficiently accelerate LWCNN through memory-oriented and computing-oriented optimizations. Firstly, a streaming architecture with hybrid CEs is designed to minimize off-chip memory access while maintaining a low cost of on-chip buffer size. Secondly, a balanced dataflow strategy is introduced for streaming architectures to enhance computational efficiency by improving efficient resource mapping and mitigating data congestion. Furthermore, a resource-aware memory and parallelism allocation methodology is proposed, based on a performance model, to achieve better performance and scalability. The proposed accelerator is evaluated on Xilinx ZC706 platform using MobileNetV2 and ShuffleNetV2. Implementation results demonstrate that the proposed accelerator can save up to 68.3% of on-chip memory size with reduced off-chip memory access compared to the reference design. It achieves an impressive performance of up to 2092.4 FPS and a state-of-the-art MAC efficiency of up to 94.58%, while maintaining a high DSP utilization of 95%, thus significantly outperforming current LWCNN accelerators.

Resistive Random-Access Memory (RRAM)-based Compute-in-Memory (CIM) architectures offer promising solutions for energy-efficient deep neural network (DNN) inference. However, conventional CIM accelerators suffer from high energy consumption due to frequent analog-to-digital (AD) and digital-to-analog (DA) signal conversions, especially when computing nonlinear activation functions (NAFs). This paper presents NeuAFG, a time-domain analog function generator designed to directly compute arbitrary NAFs in the analog domain. NeuAFG utilizes a one-hidden-layer ReLU neural network (ReLUNet) to approximate a variety of activation functions, incorporating two novel optimization algorithms: Low-Discrepancy Search (LDSearch) for efficient parameter initialization and KDE-FocusSampler for focused sampling of regions that are difficult to approximate. The hardware implementation integrates RRAM-based components and employs a Pulse-In-Pulse-Out method for robust time-domain computation. Experimental results demonstrate that NeuAFG reduces energy consumption in mainstream CIM accelerators such as ISAAC and RAELLA by 1.03× to 2.20×, with less than 1% accuracy loss in DNN inference tasks.

Large Language Models (LLMs), based on transformer architecture, have demonstrated remarkable capabilities in natural language processing tasks, enabling machines to generate human-like text and engage in meaningful dialogues. However, the exponential increase in model parameters has led to limitations in inference speed and energy efficiency. Compute-in-memory (CIM) technology offers a promising solution to accelerate AI inference by performing analog computations directly within memory, potentially reducing latency and power consumption. At the same time, CIM has been successfully applied to accelerate Convolutional Neural Networks (CNNs); however, the matrix–matrix multiplication (MatMul) operations inherent in the scaled dot-product attention of the transformer present unique challenges for direct CIM implementation. In this work, we propose InMemQK, a compute-in-memory-based attention accelerator that focuses on optimizing MatMul operations through software and hardware co-design. At the software level, InMemQK employs product quantization (PQ) to eliminate data dependencies. At the hardware level, InMemQK integrates energy-efficient time-domain MAC macros for ADC-free computations. Experimental results show InMemQK achieves 13.2×–13.9× lower power consumption than existing CIM-based accelerators.

This study addresses the critical challenge of Speech Enhancement (SE) in noisy environments, where the deployment of Deep Neural Networksolutions on microcontrollers is hindered by their extensive computational demands. Focusing on this gap, our research introduces a novel SE method optimized for MCUs, employing a 2-layer GRU model that capitalizes on perceptual speech properties and innovative training methodologies. By incorporating self-reference signals and a dual strategy of compression and recovery based on the Mel scale, we develop an efficient model tailored for low-latency applications. Our GRU-2L-128 model demonstrates a significant reduction in size and computational requirements, achieving a 14.2× decrease in model size and a 409.1× reduction in operations compared to traditional DNN methods like DCCRN, without compromising performance. This advancement offers a promising solution for enhancing speech intelligibility in resource-constrained devices, marking a pivotal step in SE research and application.

Passwords are the most widely used method of authentication and password guessing is the essential part of password cracking and password security research. The progress of deep learning technology provides a promising way to improve the efficiency of password guessing. However, current research on neural network password guessing methods mostly focuses on model structure and has overlooked the generation method. Due to the randomness of sampling, not only the generated passwords have a large number of duplicates, but also the order in which passwords generated is random, leading to inefficient password attacks. In this paper, we propose SOPG, a search-based ordered password generation method, which enables the password guessing model based on autoregressive neural network to generate passwords in approximately descending order of probability. Experiment on comparison of SOPG and Random sampling shows passwords generated by SOPG do not repeat, and when they reach the same cover rate, SOPG requires fewer inferences and far fewer generated passwords than Random sampling, which brings great efficiency improvement to subsequent password attacks. We build SOPGesGPT, a password guessing model based on GPT, using SOPG to generate passwords. Compared with the most influential models OMEN, FLA, PassGAN, VAEPass and the latest model PassGPT in one-site test, experiments show that SOPGesGPT is far ahead in terms of both effective rate and cover rate. As to cover rate that everyone recognizes, SOPGesGPT reaches 35.06%, which is 254%, 298%, 421%, 380%, 81% higher than OMEN, FLA, PassGAN, VAEPass, and PassGPT respectively.

Blockwise reconstruction with adaptive rounding helps achieve acceptable 4-bit post-training quantization accuracy. However, adaptive rounding is time intensive, and the optimization space of weight elements is constrained to a binary set, thus limiting the performance of quantized models. The optimality of block-wise reconstruction requires that subsequent network blocks remain unquantized. To address this, we propose a two-stage post-training quantization scheme, AE-Qdrop, encompassing block-wise reconstruction and global fine-tuning. In the block-wise reconstruction stage, a progressive optimization strategy is introduced as a replacement for adaptive rounding, enhancing both quantization accuracy and efficiency. Additionally, the integration of randomly weighted quantized activation helps mitigate the risk of overfitting. In the global fine-tuning stage, the weights of each quantized network block are corrected simultaneously through logit matching and feature matching. Experiments in image classification and object detection tasks validate that AE-Qdrop achieves high precision and efficient quantization. For the 2-bit MobileNetV2, AE-Qdrop outperforms Qdrop in quantization accuracy by 6.26%, and its quantization efficiency is fivefold higher.

Reverberation significantly degrades speech intelligibility, posing a substantial challenge in speech processing. While deep learning advancements offer promising solutions, current methodologies often overlook the effective integration of low-level and high-level feature representations, causing detrimental effects on overall performance. Simultaneously, prior approaches heavily rely on loss functions grounded in quantitative error metrics, which may not fully capture the perceptual intricacies of speech signals. To address these concerns, we introduce CAT-DUnet, a Unet architecture that integrates channel attention, time-frequency attention, and dilated convolution blocks to enhance feature fusion. We innovatively leverage the structural similarity as the training objective to align more closely with human perception, and investigate the effect of applying various reasonable transformations to spectrograms on the performance of the loss function. Through extensive ablation experiments, we demonstrate the effectiveness of our proposed enhancements. Our model outperforms state-of-the-art models on 6 out of 7 metrics, underscoring its exceptional performance.

Post-training quantization is pivotal in deploying convolutional neural networks for mobile applications. Block-wise reconstruction with adaptive rounding, as employed in prior works like BrecQ and Qdrop, facilitates acceptable 4-bit quantization accuracy. However, adaptive rounding is time-intensive, and its constraint on weight optimization space curtails the potential for quantization performance. The optimality of block-wise reconstruction hinges on the quantization status of subsequent network blocks. In this investigation, we delve into the theoretical underpinnings of the limitations inherent in adaptive rounding and block-wise reconstruction. Our exploration leads to the development of a post-training quantization methodology, designated as AE-Qdrop. This algorithm operates in two distinct phases: block-wise reconstruction and global fine-tuning. The block-wise reconstruction phase introduces a progressive optimization strategy, superseding adaptive rounding, which not only augments quantization precision but also significantly improves quantization efficiency. To mitigate the risk of overfitting, we introduce a random weighted quantized activation mechanism. During the global fine-tuning phase, we account for interdependencies among quantized network blocks. The weight of each network block will be corrected with logit matching and feature matching. Extensive experiments validate that AE-Qdrop achieves high-precision and efficient quantization. For instance, in the case of 2-bit MobileNetV2, AE-Qdrop outperforms Qdrop by achieving a 6.26% enhancement in quantization accuracy and quintupling the quantization efficiency.

Particle swarm optimization (PSO) is a well-known optimization algorithm that shows good performances in solving different optimization problems. However, the PSO usually suffers from slow convergence. In this article, a reinforcement-learning-based parameter adaptation method (RLAM) is developed to enhance the PSO convergence by designing a network to control the coefficients of the PSO. Moreover, based on the RLAM, a new reinforcement-learning-based PSO (RLPSO) algorithm is designed. To investigate the performance of the RLAM and RLPSO, experiments on 28 CEC 2013 benchmark functions were carried out to compare with other adaptation methods and PSO variants. The reported computational results showed that the proposed RLAM is efficient and effective and that the proposed RLPSO is superior to several state-of-the-art PSO variants.