Neurocomputing最新文献

Neural network models, celebrated for their outstanding scalability and computational capabilities, have demonstrated remarkable performance across various fields such as vision, language, and multimodality. The rapid advancements in neural networks, fueled by the deep development of Internet technology and the increasing demand for intelligent edge devices, introduce new challenges, including significant model parameter sizes and increased storage pressures. In this context, Field-Programmable Gate Arrays (FPGA) emerge as a preferred platform for accelerating neural network models, thanks to their exceptional performance, energy efficiency, and the flexibility and scalability of the system. Building FPGA-based neural network systems necessitates bridging significant differences in objectives, methods, and design spaces between model design and hardware design. This review article adopts a comprehensive analytical framework to thoroughly explore multidimensional technological implementation strategies, encompassing optimizations at the algorithmic and hardware levels, as well as compiler optimization techniques. It focuses on methods for collaborative optimization between algorithms and hardware, identifies challenges in the collaborative design process, and proposes corresponding implementation strategies and key steps. Addressing various technological dimensions, the article provides in-depth technical analysis and discussion, aiming to offer valuable insights for research on optimizing and accelerating neural network models in edge computing environments.

Nowadays, complex and expensive neural architectures are seen by many as a way to improve the performance of existing models in image recognition, voice recognition, translation, and other tasks. Such a perspective caused an increased interest in expert architecture engineering within Deep Learning. Fueled by this interest, neural architecture search originated as a promising way to automate the tedious process of constructing a deep neural network by hand. Over the last five years, we have seen an increasing number of works focusing all efforts on studying the impact of automating deep neural network design. The spotlight has recently turned from automatically discovering classification models to other more complex tasks. Motivated by a desire for high-resolution images in real-world user-centered and expert computer vision applications, architecture search for super-resolution image restoration centers in approaches capable of automatically finding efficient and well-performing models. Here, we present a survey that, beyond delving into an overview of modern approaches to automatic neural network design, focuses on the recollection and study of neural architecture search approaches that have directed their efforts at the super-resolution image restoration tasks and future lines of research found within this emerging area of study.

Obtaining fault samples for diagnosing faults in rotating machinery for engineering applications can be costly. To address this challenge, fault sample augmentation methods are used to train fault diagnosis models. However, existing techniques mainly focus on comparing augmented signal loss with real ones, overlooking the underlying vibration mechanism of rotating machinery faults. Addressing this limitation, a novel approach, called dual-loss nonlinear independent component estimation (DLNICE), is proposed to enhance understanding of fault features in vibration signals. This integrates augmentation losses in both time and frequency domains to enrich fault vibration samples. DLNICE effectively utilizes limited fault samples for augmentation by estimating nonlinear independent components, capturing key fault characteristics like impulsiveness and cyclo-stationarity. Therefore, augmented fault samples become more explainable for analyzing rotating machinery faults. Experimental evaluations on bearing and gearbox vibration samples confirm the effectiveness of DLNICE. Utilizing the augmented samples leads to an average accuracy of 86.27 % in bearing fault diagnosis, and that of 81.60 % in gearbox fault diagnosis. The results demonstrate that DLNICE excels in augmenting high-quality vibration samples of rotating machinery faults.

In recent years, AI-assisted video editing has shown promising applications. Understanding and analyzing camera language accurately is fundamental in video editing, guiding subsequent editing and production processes. However, many existing methods for camera language analysis overlook computational efficiency and deployment requirements in favor of improving classification accuracy. Consequently, they often fail to meet the demands of scenarios with limited computing power, such as mobile devices. To address this challenge, this paper proposes an efficient multi-task camera language analysis pipeline based on shared representations. This approach employs a multi-task learning architecture with hard parameter sharing, enabling different camera language classification tasks to utilize the same low-level feature extraction network, thereby implicitly learning feature representations of the footage. Subsequently, each classification sub-task independently learns the high-level semantic information corresponding to the camera language type. This method significantly reduces computational complexity and memory usage while facilitating efficient deployment on devices with limited computing power. Furthermore, to enhance performance, we introduce a dynamic task priority strategy and a conditional dataset downsampling strategy. The experimental results demonstrate that achieved a comprehensive accuracy surpassing all previous methods. Moreover, training time was reduced by 66.33%, inference cost decreased by 59.85%, and memory usage decreased by 31.95% on the 2-task dataset MovieShots; on the 4-task dataset AVE, training time was reduced by 95.34%, inference cost decreased by 97.23%, and memory usage decreased by 61.21%.

In recent years, deep learning-based image classification models have been extensively studied in academia and widely applied in industry. However, deep learning is inherently vulnerable to adversarial attacks, posing security threats to image classification models in security sensitive field, such as face recognition, medical image diagnosis and traffic sign recognition. Especially for black-box adversarial attacks, which can be carried out even without remote model information, the security issues facing deep learning are even more serious. Despite more and more attentions on this issue, existing reviews always analyze black-box adversarial attack only from one perspective, focus on only a certain application field. This paper systematically reviews and discusses existing progress, demonstrating black-box adversarial attacks from multiple perspectives and systematically classifying existing methods. Besides, we also sort out and categorize the application of current black-box adversarial attacks and identify several promising directions for future research.

Spiking Neural Networks (SNNs) exhibit the strong capability to address spatiotemporal dynamic problems. Recent research has explored the hardware SNN systems to solve the spatiotemporal problems in real-time. The Network-on-Chip (NoC) is an effective scheme for building large-scale hardware SNNs. However, for the existing NoC-based hardware SNNs, large area overhead and hardware power are consumed by their interconnections, because of complex topologies and router structures. Therefore, in this work a novel Unidirectional and Hierarchical on-Chip Interconnected Architecture (UHCIA) is proposed to address this problem. The proposed UHCIA mainly combines the novel hybrid topology of unidirectional multiple loops and rings, and uses a deflection router technique. Experimental results show that compared to other works, the UHCIA achieves $\sim$23.6X of area reduction and $\sim$6.4X of power reduction, with high system throughput and biological real-time computations.

Deep neural networks have achieved significant success across various fields, but their intrinsic black-box nature hinders the further development. Addressing the interpretability challenges, topological data analysis has emerged as a promising tool to reveal these complex models. In this work, we present a review of the emerging field of interpreting deep neural networks using topological data analysis. We organize the existing body of work into distinct analytical categories, highlighting interpretations based on the topology of data, network structural characteristics, network functional characteristics, and techniques derived from Mapper. The objective of this paper is to extract the research pattern of this area, and point out the future research direction.

Solving the unsupervised domain adaptation (UDA) task in time series is of great significance for practical applications, such as human activity recognition and machine fault diagnosis. Compared to UDA for computer vision, UDA in time series is more challenging due to the dynamics of time series data and the complex dependencies among different time steps. Existing UDA methods for time series fail to adequately capture the temporal dependencies, limiting their ability to learn domain-invariant temporal patterns. Furthermore, most UDA methods only focus on distribution adaptation on the backbone network without considering how the classifier adapts to the data distribution of the target domain. In this paper, we propose Rainforest, a three-stage UDA framework for time series. We first pre-train the backbone network through a self-supervised method called bidirectional autoregression, so that the model can comprehensively learn the temporal dependencies in time series. Next, we propose a novel meta-learning-based distribution adaptation method to achieve the joint alignment of the global and local distributions while encouraging the model to adaptively reduce the temporal dynamic differences among different domains. Finally, we design a pseudo-label-guided fine-tuning strategy to help the classifier estimate the data distribution of the target domain more accurately. Extensive experiments on four real-world time series datasets show that our Rainforest outperforms state-of-the-art methods, with an average improvement of 2.19% in accuracy and 2.41% in MF1-score.

In recent years, the deployment and operation of convolution neural networks on edge devices with limited computing capabilities have become increasingly challenging due to the large network structure and computational cost. Currently, the mainstream structured pruning algorithms mainly compress the network at the filter or layer level. However, these methods introduce too much human intervention with large granularities, which may lead to unpredictable performance after compression. In this paper, we propose a group-based automatic pruning algorithm(GAP) via kernel fusion to automatically search for the optimal pruning structure in a more fine-grained manner. Specifically, we first adopt a novel nonlinear dimensionality reduction clustering algorithm to divide the filters of each convolution layer into groups of equal size. Afterwards, we encode the mutual distribution similarity of the kernels within each group, and its KL divergence is employed as an importance indicator to determine the retained kernel groups through weighted fusion. Subsequently, we introduce an intelligent searching module that automatically explore and optimize the pruned structure of each layer. Finally, the pruned filters are permutated to form a dense group convolution and fine-tuned. Sufficient experiments show that, on two image classification datasets, for five advanced CNN models, our GAP algorithm outperforms most extant SOTA schemes, reduces artificial intervention, and enables efficient end-to-end training of compact models.

Adversarial team games (ATGs) have garnered significant attention in recent years, leading to the emergence of various solutions such as linear programming algorithms, multi-agent reinforcement learning, and game tree transformation. The ATGs involve large-scale game trees, resulting in higher time costs for computation. In this paper, we focus on expediting the solution of team-maxmin equilibrium with correlation (TMECor), which can be considered the equilibrium that maximizes the team’s payoff. To address this, we propose a transformation with seed strategies (TSS). TSS leverages reinforcement learning to calculate player strategies. We initialize the strategies of all players, referred to as seed strategies, and incorporate them into the multi-agent game tree during the transformation process. These seed strategies serve as the starting strategies for counterfactual regret minimization (CFR). CFR initializes the strategies and cumulative regret of all players based on the seed strategy. By warm starting the whole process, our method accelerates the solving of TMECor. We conducted nine experiments using Kuhn poker and Leduc Hold’em poker. The results demonstrated that TSS improved the solving speed of TMECor.