Science China Information Sciences最新文献

Operators (such as Conv and ReLU) play an important role in deep neural networks. Every neural network is composed of a series of differentiable operators. However, existing AI benchmarks mainly focus on accessing model training and inference performance of deep learning systems on specific models. To help GPU hardware find computing bottlenecks and intuitively evaluate GPU performance on specific deep learning tasks, this paper focuses on evaluating GPU performance at the operator level. We statistically analyze the information of operators on 12 representative deep learning models from six prominent AI tasks and provide an operator dataset to show the different importance of various types of operators in different networks. An operator-level benchmark, OpBench, is proposed on the basis of this dataset, allowing users to choose from a given range of models and set the input sizes according to their demands. This benchmark offers a detailed operator-level performance report for AI and hardware developers. We also evaluate four GPU models on OpBench and find that their performances differ on various types of operators and are not fully consistent with the performance metric FLOPS (floating point operations per second).

In the past decade, thanks to the powerfulness of deep-learning techniques, we have witnessed a whole new era of automated code generation. To sort out developments, we have conducted a comprehensive review of solutions to deep learning-based code generation. In this survey, we generally formalize the pipeline and procedure of code generation and categorize existing solutions according to taxonomy from perspectives of architecture, model-agnostic enhancing strategy, metrics, and tasks. In addition, we outline the challenges faced by current dominant large models and list several plausible directions for future research. We hope that this survey may provide handy guidance to understanding, utilizing, and developing deep learning-based code-generation techniques for researchers and practitioners.

The combination of contextual side information and search is a powerful paradigm in the scope of artificial intelligence. The prior knowledge enables the identification of possible solutions but may be imperfect. Contextual information can arise naturally, for example in game AI where prior knowledge is used to bias move decisions. In this work we investigate the problem of taking quantum advantage of contextual information, especially searching with prior knowledge. We propose a new generalization of Grover’s search algorithm that achieves the optimal expected success probability of finding the solution if the number of queries is fixed. Experiments on small-scale quantum circuits verify the advantage of our algorithm. Since contextual information exists widely, our method has wide applications. We take game tree search as an example.

Accurate mechanical system models are crucial for safe and stable control. Unlike linear systems, Lagrangian systems are highly nonlinear and difficult to optimize because of their unknown system model. Recent research thus used deep neural networks to generate linear models of original systems by mapping nonlinear dynamic systems into a linear space with a Koopman observable function encoder. The controller then relies on the Koopman linear model. However, without physical information constraints, ensuring control consistency between the original nonlinear system and the Koopman system is tough, as the learning process of the Koopman observation function is unsupervised. This paper thus proposes a two-stage learning algorithm that uses structural subnetworks to build a physics-informed network topology to simultaneously learn the Koopman observable functions and the system energy representation. In the Koopman matrix learning session, a quadratic-constrained optimization problem is solved to ensure that the Koopman representation satisfies the energy difference matching hard constraint. The proposed energy-preserving deep Lagrangian Koopman (EPDLK) framework effectively represents the dynamics of the Lagrangian system while ensuring control consistency. The effectiveness of EPDLK is compared with those of various Koopman observable function construction methods in multistep prediction and trajectory tracking tasks. EPDLK achieves better control consistency by guaranteeing energy difference matching, which facilitates the application of the control law generated on the Koopman system directly to the original nonlinear Lagrangian system.

In this article, a novel DDERT sensing method is proposed for large-area tactile sensing. In particular, the method utilizes a generative model to reconstruct the boundary measurement voltage of the ERT sensor into a tactile image. To improve the quality of tactile imaging, a spatial attention mechanism is incorporated into the model. Additionally, a mask constraint is introduced as prior information to ensure that the generated images contain more accurate tactile information in areas of contact with objects. Experimental results validate the proposed method is effective for the large-area robotic tactile sensing. Furthermore, the prototype of the ERT-based tactile sensor is fabricated and the sensing performance is evaluated in real robotic applications.

Waveform diversity (WD) represents a dynamic and transformative technology widely used in radar systems to enhance sensitivity and discrimination capabilities. Recently, WD techniques have been extensively explored for their potential ambiguity suppression within synthetic aperture radar (SAR) systems. Among these, the alternate transmitting mode combined with orthogonal waveforms emerges as a particularly promising solution. This study focuses on optimizing the power spectrum density (PSD) of signals to design and generate an orthogonal waveform pair that achieves both a low cross-correlation-to-autocorrelation ratio (CAR) and satisfactory imaging performance. Initially, we construct a fractional programming model with convex constraints to minimize the CAR. To address this challenge, we introduce an iterative optimization procedure for the PSD variable, which sequentially reduces the CAR. Each optimization step can be efficiently solved using a quadratically constrained quadratic program, ensuring that the resulting computational complexity remains low. Building on the optimized PSD, we established a parametric piecewise linear model to generate an orthogonal waveform pair. This model not only maintains a low CAR but achieves satisfactory imaging performance in real-time applications. Consequently, this orthogonal waveform pair effectively suppresses range ambiguity in SAR systems. Finally, we demonstrated the practicability and effectiveness of the proposed orthogonal waveforms through detailed simulation experiments, specifically targeting ambiguity suppression in conventional quad-polarization SAR systems.

This study investigates the deterministic learning (DL)-based output-feedback neural control for a class of nonlinear sampled-data systems with prescribed performance (PP). Specifically, first, a sampled-data observer is employed to estimate the unavailable system states for the Euler discretization model of the transformed system dynamics. Then, based on the observations and backstepping method, a discrete neural network (NN) controller is constructed to ensure system stability and achieve the desired tracking performance. The noncausal problem encountered during the controller deduction process is resolved using a command filter. Moreover, the regression characteristics of the NN input signals are demonstrated with the observed states. This ensures that the radial basis function NN, based on DL theory, meets the partial persistent excitation condition. Subsequently, a class of discrete linear time-varying systems is proven to be exponentially stable, achieving partial convergence of neural weights to their optimal/actual values. Consequently, accurate modeling of unknown closed-loop dynamics is achieved along the system trajectory from the output-feedback control. Finally, a knowledge-based controller is developed using the modeling results. This controller not only enhances the control performance but also ensures the PP of the tracking error. The effectiveness of the scheme is illustrated through simulation results.

Causation promotes the understanding of correlation to an advanced stage by elucidating its underlying mechanism. Although statisticians have specified the possible causal relations among correlations, inferring causal structures is impossible from only the observed correlations in the classical world. Quantum correlations encapsulating the most defining aspects of quantum physics have taken a new turn for the causal inference problem — the two-point spatial and temporal quantum correlations with observationally discernible characteristics correspond exactly to the two most basic causal structures. However, a direct causal interpretation for quantum correlations has only been established in very limited cases. Here, we explore to what extent quantum correlations promote causal inference. Theoretically, we have found that the distinguishable causal regime of two-point Pauli correlations can be expanded from a single value to an asymmetric interval, and the causal structures determining the quantum correlations can be interpreted by a simple distance criterion. Experimentally, we have devised and implemented a versatile non-unital quantum channel in an optical architecture to directly observe such an asymmetric interval. The setup enabled quantum causal inference without any requirement of active intervention, which is impossible in the classical realm. Our work facilitates the identification of causal links among quantum variables and provides insight into characterizing causation and spatial-temporal correlation in quantum mechanics.

A MoS₂ channel FeFET with a P–Si gate was developed for use as a photosensor with a memory function. A current ratio of 10⁴ was achieved at an irradiation power of 20 nW. The reliability of the device was evaluated by means of endurance tests, and a retention time of more than 1000 s was observed. Furthermore, in-sensor digital computing was verified by applying an optoelectronic hybrid logic AND gate. This novel optical sensing principle enables the development of new approaches for optoelectronic hybrid integration.

Space-surface bistatic synthetic aperture radar (SS-BiSAR) offers an additional observation angle for monostatic spaceborne SAR, making it a promising technology for high-accuracy deformation retrieval technology in local regions. Repeat-pass SS-BiSAR tomography can accurately estimate the surfaces of buildings and steep areas, effectively removing terrain phases during deformation retrieving. However, inaccuracies in the orbital ephemeris can lead to image geometry distortion, reducing image pair coherence, introducing interferometric phase errors, and consequently deteriorating tomographic precision. This paper precisely models the image geometry distortion and interferometric phase error caused by repeat-pass ephemeris error. We propose an ephemeris correction method based on the chirp-Z transform to address these issues. Furthermore, we introduce an accurate tomography model to improve 3D reconstruction accuracy. Our first SS-BiSAR tomography experiment, conducted using the Chinese Lutan-1 satellite, demonstrates that the correlation coefficient is improved by 0.16 after ephemeris error correction. Moreover, the density and precision of the tomographic point cloud are improved by 13.7% and 12.1%, respectively.