Compute-in-Memory (CIM) based on resistive random access memory (RRAM) offers significant advantages in energy efficiency and parallelism, making it a promising solution for accelerating neural networks. However, the computational accuracy, energy efficiency, and flexibility of current CIM chips are still challenged by practical issues such as device and circuit-level non-ideality and the high overhead of peripheral circuits, which remain inadequately addressed in existing designs. To address these challenges, this work proposes REF-CIM, a 40nm robust, energy efficient and flexible RRAM- CIM macro that achieves non-ideality tolerance, high energy efficiency and configurable precision, featuring: 1) a complementary multi-bit input unit (CMIU) with symmetric bit-line access; 2) a proportional current-scaling clamp circuit (PCSC); 3) a distributed tree-based sparse analog-to-digital converter (DTS-ADC); and 4) a configurable multi-mode deployment scheme for supporting diverse neural network precisions. The performance of the proposed macro is evaluated through chip measurements, considering non-ideal effects such as IR-drop, device variation, and analog circuit noise. Simulation results calibrated with measurement data demonstrate a peak energy efficiency of 29.1 TOPS/W@8bIN/8bW/16bOUT, with classification accuracy reaching 92% on the CIFAR-10 dataset under 10% device variation.

To address the performance degradation in hardware neural networks caused by non-ideal synaptic behaviors, we developed a complementary training neuromorphic architecture using self-selective memristors. The device integrates an NbOx-based volatile operating range, offering excellent endurance, with a TiOx-based non-volatile operating range, providing superior retention, thereby enabling reliable online training and inference. Comprehensive parameter evaluation, such as fluctuation, noise, endurance, retention was conducted. The proposed architecture achieved excellent pattern recognition accuracy on both the MNIST and CIFAR-10 datasets. Furthermore, stable inference performance was maintained by operating within the non-volatile regime after training.

Ag-based volatile threshold switching (TS) selectors have demonstrated considerable potential in memristor-based high-density storage due to their simple structure and excellent switching properties. However, their practical deployment is hindered by poor endurance and switching uniformity which are caused by excessive accumulation and random distribution of Ag ions during cycling. In this work, we introduce Ti insertion layer and post-deposition annealing engineering to improve the selector performance. Selectors with different structures and fabrication processes are designed and analyzed based on electrical characteristics, and the switching mechanisms are qualitatively elaborated. The results show that the selector with 3nm Ti layer and 400°C annealing could effectively reduce Ag accumulation and enables controlled filament formation, resulting in improved endurance exceeding 109 cycles and device-to-device variation of less than 10%.

In this work, we propose a highly accurate and energy-efficient on-chip learning algorithm, which leverages the intrinsic variation in analog RRAM to overcome the non-convergence in conventional Manhattan weight update rules due to the limited intermediate RRAM states. The proposed learning approach achieves approximately a 15x improvement in classification accuracy despite the limited number (similar to 20) of intermediate states available in RRAM cells. Our findings underscore the potential for reliable and efficient on-chip learning in RRAM-based neural network systems.

To address the issue of near-field localization within the spherical wavefronts, in this paper, a cumulant matrices-based algorithm is proposed based on the exact spatial propagation model. Multiple fourth-order cumulant matrices are constructed with the changed position number of the array elements, which are then cascaded to a cumulant-based long matrix. Next, the covariance matrix of the cumulant matrix is computed, and eigen-decomposition is performed to estimate the corresponding array manifold matrix by using the ESPRIT. The subsequent step involves extracting amplitude attenuation and spatial phase from the estimated manifold matrix. The least squares method is then employed to yield unambiguous coarse estimates and ambiguity-resolving fine estimates for angle and range parameters. In addition, the asymptotic variance for the proposed estimator is theoretically derived. Simulation results demonstrate that the correctness of the proposed algorithm.

White balance (WB) is a critical back-end processing function in image sensors to keep color constancy under various lighting conditions by adjusting the RGB color channel gain values. In this work, we proposed a novel gate-controlled gain-tuning method for a fully depleted silicon-on-insulator (FDSOI) one-transistor (1 T) pixel to achieve WB inside the sensor. Based on the pixel structure of a p-well under the buried oxide for sensing and an n-type nMOSFET on the top for readout, the 1 T pixel output gain is modulated by the MOSFET gate according to the transistor transfer characteristics. About 5x gain modulation range in RGB spectrum photoresponse (nonlinearity < 3%) is experimentally demonstrated in the devices fabricated by 22 nm FDSOI-based technology. The scheme for in-sensor WB demonstration is provided with a novel 1 T pixel array design, and the evaluation result shows in-sensor WB achieving an almost equivalent performance (Delta-E deviation < 1) compared with using conventional back-end WB. (c) 2024 The Japan Society of Applied Physics

Transient drain current (Ids) overshoot is generally observed during the switching process of low-temperature poly-silicon thin-film transistors (LTPS-TFTs), followed by a gradual recovery to the steady-state afterwards. To model this behavior and to predict its impact at the circuit level, a surface-potential based LTPS-TFT compact model is established first, considering the trap occupation in poly-silicon (poly-Si) channel and the gate oxide. Then a time-dependent trapping and detrapping model is brought forward to describe the Ids transient characteristics of LTPS-TFTs. The model can be used in circuit simulation to optimize the circuits in display.

Recently, many deep convolutional dictionary learning-based methods, integrating the traditional image representation methods with deep neural networks, have achieved great success in various image processing tasks. However, the existing approaches can be further improved with the following considerations: (1) They congenitally suffer from the high cross-channel correlation loss for color image processing tasks since they usually treat each color channel independently, not in a whole perspective. (2) They only build up a single reconstruction dictionary learning model to directly approximate images using several single dictionary atoms, which cannot make full use of the representative ability of the model. In this paper, we propose a simultaneously learning deep quaternion reconstruction and noise convolutional dictionary model. To fully explore the cross-channel correlation, we use the quaternion method to process the color image in a holistic way. An adaptive attentional weight of reconstruction and noise learning module is also developed for the optimal combination between reconstruction and noise learning. Experimental results for synthesis and real color image denoising have demonstrated the superiority of the proposed method over other state-of-the-art methods.

In this work, we proposed and fabricated a novel artificial nociceptor based on short-term memory (STM) TFT with fully BEOL-compatible process. The proposed device well emulates four essential characteristics of nociceptors and exhibits adjustable sensitization level controlled by gate voltage. We further demonstrated a pain perception system composed of the nociceptor and a leaky integrate-and-fire (LIF) neuron.

Motivated by the advance of deep learning methods, deep unfolding methods such as deep convolutional dictionary learning have achieved great success in image denoising tasks. The main advantages are inheriting both the merits of deep learning (strong learning capacity) and traditional machine learning (powerful interpretable capacity). We observe that the update of dictionaries and coefficients is highly correlated with the previous iterative stage information for deep unfolding-based methods. However, most existing deep convolutional dictionary learning methods deal with each iteration step individually, ignoring the inner-memory within the stage and cross-memory across the stages. To alleviate these issues, we propose a dynamic inner-cross memory augmented attentional dictionary learning (M2ADL) network with attention guided residual connection module, which utilizes the previous important stage features such that better uncovering the inner-cross information. Specifically, the proposed inner-cross memory fully utilizes the previous stage's hidden and last-layer information to learn the dictionary. In addition, we develop a dual attention-guided residual connection module to well exploit the deep feature learning ability to capture the spatial-spectral attention across the deep tensor-based features. Considerable experiments on both synthetic and real image datasets demonstrate the superiority of the proposed method over other state-of-the-art methods.

In this work, we investigate the physical mechanisms of endurance degradation of SiOx:Ag-based memristors with electrical and material characterizations. The results indicate that the density of Si dangling bonds will influence the diffusion direction of Ag atoms during the relaxation process. High density of Si dangling bonds lead to Ag diffusion to the SiOx layer and endurance degradation, which is supported by the temperature and frequency dependent conduction mechanism. Finally, excellent endurance above 10(7) cycles is achieved by using oxygen annealing according to the mechanism.

By integrating the storage and computing functions on the fundamental elements, computing in-memory (CIM) technology is widely considered as a novel computational paradigm that can break the bottleneck of Von Neumann architecture. Nonvolatile memory device is an appropriate hardware implementation approach of CIM, which possess significantly advantages, such as excellent scalability, low consumption, and versatility. In this paper, first we introduce the basic concept of CIM, including the technical background and technical characteristics. Then, we review the traditional and novel nonvolatile memory devices, flash and resistive random access memory (RRAM), used in non-volatile based computing in-memory (nvCIM) system. After that, we explain the operation modes of nvCIM: in-memory analog computing and in-memory digital computing. In addition, the applications of nvCIM are also discussed, including deep learning accelerator, neuromorphic computing, and stateful logic. Finally, we summarize the current research advances in nvCIM and provide an outlook on possible research directions in the future.

Existing local descriptors are designed for gray-scale images to insufficiently consider the color information and hardly extract the discriminative features. This work first proposes a Michelson local descriptor (MLD) to effectively extract local features and a cross-channel similarity (CCS) measure to capture high-order correlations between color channels. Based on the proposed CCS measure, we devise a general CCS-based local descriptor (CCS-LD) framework for color images. Using this simple yet effective framework, most local descriptors for gray-scale images can be easily extended as new CCS-based local descriptors for color images. As examples, integrating the proposed MLD, local binary pattern, histogram of oriented gradients, and scale-invariant feature transform into the proposed framework, we obtain the CCS-based local descriptors for color images, respectively. Extensive experiments for texture classification, face recognition, and kinship verification have demonstrated the superior performance of the proposed methods over other state-of-the-art methods.

Most engineering design optimization problems are complex and expensive multi-objective optimization problems with multiple constraints. This paper proposes an improved surrogate-based multi-objective optimization algorithm (SBMO) using an adaptive weight vector generation method to address them. The main idea of the improved SBMO is to update the weight vectors of sub-problems adaptively according to the shape of the current Pareto front (PF) at each iteration while using SBMO to obtain Pareto optimal solutions. First, the improved SBMO decomposes a multi-objective optimization problem into a set of single-objective optimization sub-problems and builds surrogate models for each objective. Second, solutions are obtained by solving the acquisition problems for the sub-problems under the infill-sampling criteria. Third, all the solutions obtained will be evaluated and used to update the surrogate models to share the search information. At each iteration, the improved SBMO will divide the topology of the current PF evenly and select random points between the segment points. Well-distributed weight vectors will be generated based on the random points. This weight vector generation method can significantly improve the distribution of Pareto optimal solutions. The studies on benchmark test instances and aerodynamic design optimization of an airfoil indicate that the improved SBMO can obtain Pareto optimal solutions with better distribution than SBMO in a small number of sample points, and offers great potential to solve an expensive multi-objective optimization problem.

This work addresses the issue of spectrum sensing with random arrival and departure of primary signals. We first design a convolutional neural network (CNN) with outputs as the posterior probabilities of the arrival and departure of primary signals, leading to a CNN-based detector with the ratio of the posterior probabilities (i.e., the outputs of the CNN) as a test statistic. To further enhance the attention of the network on the switch feature of channel states, we design a switch attention module (SAM) that adaptively weights the received signals. Replacing the convolution plus maximum pooling block in the CNN detector with the SAM block leads to an SAM-CNN detector. Simulations show that the proposed CNN detector significantly outperforms existing detectors, and further improvement of detection probability by 19% is achieved by the SAM-CNN detector.

Comprehensive geriatric assessment (CGA)-based cardiac rehabilitation (CR) program is essential for patients before and after transcatheter aortic valve implantation (TAVI). This study aimed to explore the values of CGA and exercise capacity in CR for patients referred to TAVI. A retrospective analysis was conducted in 90 patients referred to TAVI from January to October 2019. CR strategies started before TAVI. The association between clinical characteristics, CGA, and change in six-minute walk distance (Δ6MWD) was analyzed with multivariate regression models. Most of patients had cognitive impairment (50%), malnutrition (61%), and frailty (83%). After the CR, the proportion of cognitive impairment, malnutrition, and frail patients was significantly decreased by 21%, 40%, and 57%, respectively (p = 0.002, p <0.001, p <0.001). The 6MWD at a month after discharge (291.9 ± 98.8 m) was significantly improved than that at discharge after TAVI (218.8 ± 114.3m, p <0.001). The multivariate regression analysis indicated body mass index (BMI; Δ6MWD:12.0, 95% confidence interval [CI] 0.3 to 23.8, p = 0.045), frailty (Δ6MWD: -57.9, 95% CI -81.8 to -34.1, p <0.001) and malnutrition (Δ6MWD: -25.1, 95% CI -47.0 to -3.2, p = 0.026) as the associated predictors of Δ6MWD. In conclusion, functional status in patients referred to TAVI could be improved by CGA-based CR. BMI, frailty, and malnutrition were associated with the efficacy of CR on exercise capacity. CGA can play the important role in the evaluation and making strategies for CR in patients.

This paper proposes a half-selection stability fault, which belongs to a type of dynamic fault that cannot be covered by traditional March algorithms. As the voltage decrease, the impact of the fault gradually increases and eventually causes the memory to fail. In order to solve the problem, March_HS algorithm is proposed. The new algorithm not only can effectively test half-selection stability faults, but also has a high fault coverage. An improved random fault injection method is proposed to verify the fault coverage of the March_HS. This method can simulate the faults of SRAM more realistically during the actual chip manufacturing process. The algorithm and random fault injection are verified on a low power MCU, based on TSMC 28nm. The experimental results show that random fault injection realizes location, resistance and number of injected faults are all random. At 0.8V, the fault coverage of March_HS can be improved by 25.7%.

The magnetization orientation dependence of Gilbert damping and ultrafast demagnetization were investigated in single-crystalline ${\mathrm{Co}}_{50}{\mathrm{Fe}}_{50}$ films grown on a MgO(100) substrate by the time-resolved magneto-optical Kerr effect technique. The intrinsic Gilbert damping coefficient extracted from the time evolution of magnetization precessions shows a remarkably large anisotropy with a ratio of more than 300% for magnetization orientations along the $\ensuremath{\langle}100\ensuremath{\rangle}$ and $\ensuremath{\langle}110\ensuremath{\rangle}$ axes. Such a large anisotropy of Gilbert damping persists to high frequencies up to $50\phantom{\rule{4pt}{0ex}}\mathrm{GHz}$, where the effect of two-magnon scattering is suppressed. In contrast to the anisotropic Gilbert damping, the ultrafast demagnetization time and ratio for different in-plane magnetization orientations are nearly isotropic with the magnetization orientation. Although anisotropic Gilbert damping can be explained by the variation of spin-orbit coupling with the magnetization orientations, our results demonstrate that the role of spin-orbit coupling in the high nonequilibrium state after ultrafast laser excitation is isotropic in driving ultrafast spin-flip processes.

Artificially-hydrophobized soil has been used in geotechnical applications, such as slope stabilization. This application requires a thorough understanding of the soil's hydromechanical properties to inform stability assessment. The shearing behavior of dry and fully saturated hydrophobized sand has been extensively investigated in the literature. Yet, knowledge on the effects of unsaturation on the soil's hydmmechanical properties is scarce. Thus, how hydrophobic coating affects geotechnical properties such as compressibility and stress-dilatancy relation is unclear. This study conducted a comprehensive and systematic test program to, for the first time, quantify the hydromechanical properties of unsaturated hydrophobized Toyoura sand under wide ranges of confining pressure (25-300 kPa) and degrees of saturation (S; 0%-100%). The sand was hydrophobized by dichlorodimethylsilane (DMDCS). Constant-water-content shearing tests and compression tests were performed using the direct-shear box and oedometer apparatus, respectively. The test results revealed that the hydrophobic coating (i) made the compressibility index, swelling index, and peak friction angle independent of S; (ii) switched the sand's stress-strain behavior from strain-softening to strain-hardening at any degree of saturation; (iii) made the low-stress nonlinearity associated with sand dilatancy almost vanish; and (iv) made the peak friction angle to be practically independent of stress.