Science China Information Sciences最新文献

The logical blocking of finite state machines (FSMs) is examined at the three levels of formulation, detection, and search from an STP viewpoint (semi-tensor product of matrices). The research idea regards an FSM as a logical system. The realizing method treats the event sequence exciting an FSM as the input signal of a logical system and treats the current states of an FSM as the states of a logical system. Based on a recently developed bilinear dynamic model of FSMs, a difference equation-like model is first proposed to describe the logical blocking. By defining a loop structure of FSMs and using the difference equation-like model, a criterion is built by which whether a given FSM is blocking can be easily judged. If it is, several algorithms are designed to find all the logical blocking of the FSM. Further, these results are extended to apply to the case of nondeterministic FSMs and, thus, to networks of FSMs. The proposed STP approach may provide a new angle for considering the problems of FSMs, and the presented results may strengthen the links between systems governed by human-designed rules and systems governed by natural laws.

Today’s deep learning models face an increasing demand to handle dynamic shape tensors and computation whose shape information remains unknown at compile time and varies in a nearly infinite range at runtime. This shape dynamism brings tremendous challenges for existing compilation pipelines designed for static models which optimize tensor programs relying on exact shape values. This paper presents TSCompiler, an end-to-end compilation framework for dynamic shape models. TSCompiler first proposes a symbolic shape propagation algorithm to recover symbolic shape information at compile time to enable subsequent optimizations. TSCompiler then partitions the shape-annotated computation graph into multiple subgraphs and fine-tunes the backbone operators from the subgraph within a hardware-aligned search space to find a collection of high-performance schedules. TSCompiler can propagate the explored backbone schedule to other fusion groups within the same subgraph to generate a set of parameterized tensor programs for fused cases based on dependence analysis. At runtime, TSCompiler utilizes an occupancy-targeted cost model to select from pre-compiled tensor programs for varied tensor shapes. Extensive evaluations show that TSCompiler can achieve state-of-the-art speedups for dynamic shape models. For example, we can improve kernel efficiency by up to 3.97× on NVIDIA RTX3090, and 10.30 × on NVIDIA A100 and achieve up to five orders of magnitude speedups on end-to-end latency.

In this paper, we focus on the state and parameter identification problem of a hydrodynamical system. This system is modeled as a linearized water wave equation (LWWE), a hyperbolic state-space model coupled with a Laplace equation. We assume that the wave elevation at two distinct points is the only measurement of water waves. We show that the state and water depth can be reconstructed from this point measurement records. The identification problem is recast as an optimization problem over an infinite-dimensional space. We propose the adjoint method-based identification algorithm to generate an estimated state and water depth. We then performed a numerical simulation to show the effectiveness of our designed algorithm by comparing it with existing studies.

In the wake of rapid advancements in artificial intelligence (AI), we stand on the brink of a transformative leap in data systems. The imminent fusion of AI and DB (AI×DB) promises a new generation of data systems, which will relieve the burden on end-users across all industry sectors by featuring AI-enhanced functionalities, such as personalized and automated in-database AI-powered analytics, and self-driving capabilities for improved system performance. In this paper, we explore the evolution of data systems with a focus on deepening the fusion of AI and DB. We present NeurDB, an AI-powered autonomous data system designed to fully embrace AI design in each major system component and provide in-database AI-powered analytics. We outline the conceptual and architectural overview of NeurDB, discuss its design choices and key components, and report its current development and future plan.

To advance deep learning methodologies in the next decade, a theoretical framework for reasoning about modern neural networks is needed. While efforts are increasing toward demystifying why deep learning is so effective, a comprehensive picture remains lacking, suggesting that a better theory is possible. We argue that a future deep learning theory should inherit three characteristics: a hierarchically structured network architecture, parameters iteratively optimized using stochastic gradient-based methods, and information from the data that evolves compressively. As an instantiation, we integrate these characteristics into a graphical model called neurashed. This model effectively explains some common empirical patterns in deep learning. In particular, neurashed enables insights into implicit regularization, information bottleneck, and local elasticity. Finally, we discuss how neurashed can guide the development of deep learning theories.

The conventional reconfigurable intelligent surface (RIS) is limited to reflecting incident signals, thereby imposing constraints on the placement of the transmitter and receiver, which hinders achieving comprehensive signal coverage across an entire area. This paper investigates a simultaneously transmitting and reflecting (STAR)-RIS-aided simultaneous wireless information and power transfer (SWIPT) system with a nonlinear energy harvesting model under three different RIS transmission protocols: energy splitting (ES), time switching (TS), and mode switching (MS). The objective of this paper is to maximize the weighted sum power (WSP) of all energy harvesting receivers (EHRs) while ensuring fairness in the collected power among them. This is achieved by jointly optimizing the transmit beamforming at the base station (BS) and the transmission and reflection coefficients at the STAR-RIS, subject to rate constraints for information decoding receivers (IDRs), transmit power constraint at the BS, and coefficient constraints of each element at the STAR-RIS corresponding to the three protocols. Solving this optimization problem poses challenges because of the complicated objective function and numerous coupled optimization variables of the ES STAR-RIS. To address this complexity, an efficient alternating optimization (AO) approach is proposed as an iterative solution method that achieves suboptimal results. The AO algorithm is then extended to MS STAR-RIS and TS STAR-RIS. Specifically, for MS STRA-RIS, binary constraints in the STAR-RIS coefficient optimization subproblem are handled using the first-order approximation technique along with the penalty function method. For TS STAR-RIS, apart from optimizing BS transmit beamforming and STAR-RIS coefficients subproblems, the transmission and reflection time allocation of STAR-RIS also needs optimization. Numerical findings demonstrate that compared to conventional RIS-aided systems, utilizing three different protocols in a STAR-RIS-aided system can enhance power collection at EHRs while expanding the receiver placement range. Furthermore, TS STAR-RIS performs best when the IDRs do not require high achieved rates. Otherwise, ES is the best choice.

The realistic security of quantum key distribution (QKD) systems is currently a hot research topic in the field of quantum communications. There are always defects in practical devices, and eavesdroppers can make use of the security risk points of various devices to obtain key information. To date, current types of security analysis tend to analyze each security risk point individually, thereby posing great challenges for the overall security evaluation of QKD systems. In this paper, for the first time, we employ machine learning algorithms to identify the defects of different devices and certain attacks in real time, with an accuracy of 98%. It provides a novel solution for the practical security evaluation of QKD systems, thereby addressing the bottleneck problem of multiple risk points being difficult to address simultaneously in QKD systems, thus paving the way for the future large-scale application of quantum communication networks.

Inspired by biological visual systems, optoelectronic synapses with image perception, memory retention, and preprocessing capabilities offer a promising pathway for developing high-performance artificial perceptual vision computing systems. Among these, oxide-based optoelectronic synaptic transistors are well-known for their enduring photoconductive properties and ease of integration, which hold substantial potential in this regard. In this study, we utilized indium gallium zinc oxide as a semiconductor layer and high-k ZrAlO_x as a gate dielectric layer to engineer low-power high-performance synaptic transistors with photonic memory. Crucial biological synaptic functions, including excitatory postsynaptic currents, paired-pulse facilitation, and the transition from short-term to long-term plasticity, were replicated via optical pulse modulation. This simulation was sustained even at an operating voltage as low as 0.0001 V, exhibiting a conspicuous photonic synaptic response with energy consumption as low as 0.0845 fJ per synaptic event. Furthermore, an optoelectronic synaptic device was employed to model “learn-forget-relearn” behavior similar to that exhibited by the human brain, as well as Morse code encoding. Finally, a 3 × 3 device array was constructed to demonstrate its advantages in image recognition and storage. This study provides an effective strategy for developing readily integrable, ultralow-power optoelectronic synapses with substantial potential in the domains of morphological visual systems, biomimetic robotics, and artificial intelligence.

Real-time systems are widely implemented in the Internet of Things (IoT) and safety-critical systems, both of which have generated enormous social value. Aiming at the classic schedulability analysis problem in real-time systems, we proposed an exact Boolean analysis based on interference (EBAI) for schedulability analysis in real-time systems. EBAI is based on worst-case interference time (WCIT), which considers both the release jitter and blocking time of the task. We improved the efficiency of the three existing tests and provided a comprehensive summary of related research results in the field. Abundant experiments were conducted to compare EBAI with other related results. Our evaluation showed that in certain cases, the runtime gain achieved using our analysis method may exceed 73% compared to the state-of-the-art schedulability test. Furthermore, the benefits obtained from our tests grew with the number of tasks, reaching a level suitable for practical application. EBAI is oriented to the five-tuple real-time task model with stronger expression ability and possesses a low runtime overhead. These characteristics make it applicable in various real-time systems such as spacecraft, autonomous vehicles, industrial robots, and traffic command systems.

Current knowledge distillation (KD) methods primarily focus on transferring various structured knowledge and designing corresponding optimization goals to encourage the student network to imitate the output of the teacher network. However, introducing too many additional optimization objectives may lead to unstable training, such as gradient conflicts. Moreover, these methods ignored the guidelines of relative learning difficulty between the teacher and student networks. Inspired by human cognitive science, in this paper, we redefine knowledge from a new perspective — the student and teacher networks’ relative difficulty of samples, and propose a pixel-level KD paradigm for semantic segmentation named relative difficulty distillation (RDD). We propose a two-stage RDD framework: teacher-full evaluated RDD (TFE-RDD) and teacher-student evaluated RDD (TSE-RDD). RDD allows the teacher network to provide effective guidance on learning focus without additional optimization goals, thus avoiding adjusting learning weights for multiple losses. Extensive experimental evaluations using a general distillation loss function on popular datasets such as Cityscapes, CamVid, Pascal VOC, and ADE20k demonstrate the effectiveness of RDD against state-of-the-art KD methods. Additionally, our research showcases that RDD can integrate with existing KD methods to improve their upper performance bound. Codes are available at https://github.com/sunyueue/RDD.git.