Neural Network Simulation Research Articles

Controlled Migration of Lithium Cations by Diamine Bridges in Water-Processable Polymer-Based Solid-State Electrolyte Memory Layers for Organic Synaptic Transistors.

Synaptic transistors require sufficient retention (memory) performances of current signals to exactly mimic biological synapses. Ion migration has been proposed to achieve high retention characteristics but less attention has been paid to polymer-based solid-state electrolytes (SSEs) for organic synaptic transistors (OSTRs). Here, OSTRs with water-processable polymer-based SSEs, featuring ion migration-controllable molecular bridges, which are prepared by reactions of poly(4-styrenesulfonic acid) (PSSA), diethylenetriamine (DETA), and lithium hydroxide (LiOH) are demonstrated. The ion conductivity of PSSA:LiOH:DETA (1:0.4:X, PLiD) films is remarkably changed by the molar ratio (X) of DETA, which is attributed to the extended distances between the PSSA chains by the DETA bridges. The devices with the PLiD layers deliver noticeably changed hysteresis reaching an optimum at X = 0.2, leading to the longest retention of current signals upon single/double pulses. The long-term potentiation test confirms that the present OSTRs can gradually build up the postsynaptic current by gate pulses of -2V, while the long-term depression can be adjusted by varying the depression gate pulses (≈0.2-1.2V). The artificial neural network simulations disclose that the present OSTRs with the ion migration-controlled PLiD layers can perform synaptic processes with an accuracy of ≈96%.

Preserving potato perfection: Optimizing innovative drying techniques for maintaining physicochemical attributes and starch structure

To address the limitations of conventional hot air drying (HAD) for potato dehydration, pulsed vacuum drying (PVD) and humidity-controlled HAD (HC-HAD) were proposed as innovative techniques. In PVD, varying vacuum pulse durations (6, 8, 10, 12, and 14 min) and atmospheric pulse durations (2, 3, 4, and 5 min) were investigated. Under HC-HAD, different humidity control scenarios (40,30+30,90+CF, 50,30+30,90+CF, 40,30+CF, and 50,30+CF) were examined. The study revealed that a 14:2 duration pulsing in PVD retained the color and achieved the highest total phenol content (TPC) and antioxidation activity (conserving 97.50 % and 84.67 % of their fresh values, respectively) but had the longest drying time (260.00 min) compared to HC-HAD and HAD (139.45 min). The 50,30+CF strategy of HC-HAD exhibited the quickest drying time (116.79 min), maintained the color, TPC (87.97 % of the fresh value), and antioxidation activity (71.40 % of the fresh value), and achieved the best potato microstructure and starch structure. Furthermore, Artificial Neural Network simulations for HC-HAD conditions demonstrated significant reliability. These findings provide valuable insights for potato researchers and producers seeking to enhance preservation techniques and prolong the shelf life of potatoes as a staple crop.

Engineering of TiN/ZnO/SnO2/ZnO/Pt multilayer memristor with advanced electronic synapses and analog switching for neuromorphic computing

The two-terminal memristor is a promising neuromorphic artificial electronic device, mirroring biological synapses in structure and replicating various synaptic functions. Despite its advantages, challenges in achieving high reliability, gradual switching, and low energy consumption hinder progress in neuromorphic devices. This study explores electronic synapses and simulates analog switching in a Pt/TiN/ZnO/SnO2/ZnO/Pt multilayer (ML) configuration, featuring a ∼3 nm SnO2 layer between ZnO layers. Results show enhanced cycling endurance (more than 250 cycles), resistance window (102), tunable synaptic plasticity, and multilevel switching. ML memristors exhibit low coefficient of variation (4.5 %) in set voltage, low energy consumption (set = ∼0.12 nj, reset = ∼0.1 nj), and fast switching speeds (set = 300 ns, reset = 200 ns), suitable for high-density memory and neuromorphic systems. They successfully emulate synaptic functions, including paired-pulse facilitation (PPF), spike voltage-dependent plasticity (SVDP), spike width-dependent plasticity (SWDP), spike frequency-dependent plasticity (SFDP), and post-tetanic potentiation (PTP). Modulating pulse amplitude and width achieves multilevel conductance in long-term potentiation (LTP) and long-term depression (LTD). Using nonlinear conductance data, a 96.5 % image pattern recognition accuracy is achieved in a deconvolution neural network (DNN) simulation. These results highlight the ML memristor's potential in efficient neuromorphic computing systems.

Physics-informed neural network simulation of thermal cavity flow

Physics-informed neural networks (PINNs) are an emerging technology that can be used both in place of and in conjunction with conventional simulation methods. In this paper, we used PINNs to perform a forward simulation without leveraging known data. Our simulation was of a 2D natural convection-driven cavity using the vorticity-stream function formulation of the Navier-Stokes equations. We used both 2D simulations across the x and z domains at constant Rayleigh (Ra) numbers and 3D simulations across the x, z and Ra domains. The 3D simulation was tested for a PINN’s ability to learn solutions in a higher-dimensional space than standard simulations. The results were validated against published solutions at Ra values of 103\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$^{3}$$\\end{document}, 104\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$^{4}$$\\end{document}, 105\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$^{5}$$\\end{document}, and 106\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$^{6}$$\\end{document}. Both the 2D simulations and 3D simulations successfully matched the expected results. For the 2D cases, more training iterations were needed for the model to converge at higher Ra values (105\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$^5$$\\end{document} and 106\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$^6$$\\end{document}) than at lower Ra (103\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$^3$$\\end{document} and 104\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$^4$$\\end{document}) indicating increased nonlinear fluid-thermal coupling. The 3D case was also able to converge but, but it required more training than any of the 2D cases due to the curse of dimensionality. These results showed the validity of standard simulations via PINNs and the feasibility of higher-order parameter space solutions that are not possible using conventional methods. They also showcased the additional computational demand associated with increasing the dimensionality of the learned parameter space.

Lasing‐Assisted Synthesis of Metal–Organic Frameworks (MOFs) and Its Application to Memory and Neuromorphic Devices

AbstractRecently, metal–organic frameworks (MOFs) have gained attention in the field of electronics owing to their capability to tune their electrical characteristics. However, conventional methods for synthesizing MOFs pose challenges for their integration into electronic devices because of their long synthesis times and complex transfer steps. In this study, for the first time, lasing‐assisted synthesis (LAS) is used to rapidly and directly synthesize MOFs. These are applied to resistive random access memory (RRAM) devices. Using the LAS method, Cu(BDC) and Cu(BTC) are synthesized in a remarkably short time (≈5 min) and formed directly on metal substrates as thin films. This simplified their integration into RRAMs. The Cu(BDC)‐ and Cu(BTC)‐based RRAMs are evaluated for their potential in memory and neuromorphic applications. Both devices demonstrated nonvolatile memory capabilities with a remarkable data retention time of 104 s and long‐term plasticity (LTP) in response to voltage stimuli. However, the suitability of each device for a specific application varies depending on the type of MOFs used. The Cu(BTC)‐based RRAM is more suitable for memory applications because of its higher on/off ratio, longer endurance, and more data storage capacity. Conversely, Cu(BDC)‐based RRAM is highly effective in neural network simulation, achieving higher classification accuracy.

Nondifferentiable activity in the brain.

Spike raster plots of numerous neurons show vertical stripes, indicating that neurons exhibit synchronous activity in the brain. We seek to determine whether these coherent dynamics are caused by smooth brainwave activity or by something else. By analyzing biological data, we find that their cross-correlograms exhibit not only slow undulation but also a cusp at the origin, in addition to possible signs of monosynaptic connectivity. Here we show that undulation emerges if neurons are subject to smooth brainwave oscillations while a cusp results from nondifferentiable fluctuations. While modern analysis methods have achieved good connectivity estimation by adapting the models to slow undulation, they still make false inferences due to the cusp. We devise a new analysis method that may solve both problems. We also demonstrate that oscillations and nondifferentiable fluctuations may emerge in simulations of large-scale neural networks.

Three-Dimensional Physics-Informed Neural Network Simulation in Coronary Artery Trees

Three-Dimensional Physics-Informed Neural Network Simulation in Coronary Artery Trees

Efficient modeling of liquid splashing via graph neural networks with adaptive filter and aggregator fusion

PurposeThe purpose of this study is to advance the computational modeling of liquid splashing dynamics, while balancing simulation accuracy and computational efficiency, a duality often compromised in high-fidelity fluid dynamics simulations.Design/methodology/approachThis study introduces the fluid efficient graph neural network simulator (FEGNS), an innovative framework that integrates an adaptive filtering layer and aggregator fusion strategy within a graph neural network architecture. FEGNS is designed to directly learn from extensive liquid splash data sets, capturing the intricate dynamics and intrinsically complex interactions.FindingsFEGNS achieves a remarkable 30.3% improvement in simulation accuracy over traditional methods, coupled with a 51.6% enhancement in computational speed. It exhibits robust generalization capabilities across diverse materials, enabling realistic simulations of droplet effects. Comparative analyses and empirical validations demonstrate FEGNS’s superior performance against existing benchmark models.Originality/valueThe originality of FEGNS lies in its adaptive filtering layer, which independently adjusts filtering weights per node, and a novel aggregator fusion strategy that enriches the network’s expressive power by combining multiple aggregation functions. To facilitate further research and practical deployment, the FEGNS model has been made accessible on GitHub (https://github.com/nanjinyao/FEGNS/tree/main).

A Reconfigurable All-Optical-Controlled Synaptic Device for Neuromorphic Computing Applications.

Retina-inspired visual sensors play a crucial role in the realization of neuromorphic visual systems. Nevertheless, significant obstacles persist in the pursuit of achieving bidirectional synaptic behavior and attaining high performance in the context of photostimulation. In this study, we propose a reconfigurable all-optical controlled synaptic device based on the IGZO/SnO/SnS heterostructure, which integrates sensing, storage and processing functions. Relying on the simple heterojunction stack structure and the role of energy band engineering, synaptic excitatory and inhibitory behaviors can be observed under the light stimulation of ultraviolet (266 nm) and visible light (405, 520 and 658 nm) without additional voltage modulation. In particular, junction field-effect transistors based on the IGZO/SnO/SnS heterostructure were fabricated to elucidate the underlying bidirectional photoresponse mechanism. In addition to optical signal processing, an artificial neural network simulator based on the optoelectrical synapse was trained and recognized handwritten numerals with a recognition rate of 91%. Furthermore, we prepared an 8 × 8 optoelectrical synaptic array and successfully demonstrated the process of perception and memory for image recognition in the human brain, as well as simulated the situation of damage to the retina by ultraviolet light. This work provides an effective strategy for the development of high-performance all-optical controlled optoelectronic synapses and a practical approach to the design of multifunctional artificial neural vision systems.

Liquid fuel production from syngas: Simulation and optimization using artificial neural network

This study aimed to increase liquid fuel (C5+) product selectivity from CO and H2 conversion over a Cu/Fe-based catalyst and followed by artificial neural network simulation to enable a profound visualization strategy for Fischer–Tropsch (FT) process. The experiments were performed in a fixed-bed reactor over temperatures (T) of 483–553 K, pressures (P) of 14.5–203 psig, space velocities (SV) of 1000–2900 hr−1, and H₂/CO feed ratios of 1–6.2. Mathematical modeling (MM) and artificial neural network (ANN) were utilized to simulate the nonlinear CO conversion and product selectivity of the FT process. Nine simultaneous reaction pathway was developed to evaluate the reactant and product behaviors under varying FT conditions. In addition, a multilayered parameter processed by a neural network was used to predict the performance of the CO hydrogenation process. A comparison between these two approaches demonstrated that ANN more closely mimicked the nonlinear conversion and selectivity of the FT process than MM, and the correlation coefficients for the predicted mole fractions and CO conversion were 0.991 and 0.986, respectively. A hybrid Genetic Algorithm–ANN procedure was constructed to predict the optimum operating conditions to increase liquid fuel over the catalyst. The optimum C5+ selectivity (75.1 %), which meets the FT process objective, was achieved under operating conditions of T = 483 K, P = 203 psig, SV = 1303 hr−1, and H₂/CO feed ratio = 1.01 for process design optimization.

A neural network simulator for educational purposes

Artificial neural networks are inspired by biologic processes. Artificial neural networks are important because they can be used to deduct a function from observations, in other words artificial neural networks can learn from experience. Artificial neural network simulator to fulfill a need into the growing interest of neural network education is introduced in this study. NeuroQuick Laboratory simulator is implemented using object-oriented programming by Delphi programming and these classes can be used to create a standalone application with artificial neural networks. The NeuroQuick Laboratory Simulator is designed for a broad range of users, including beginning graduate/advanced undergraduate students, engineers, and scientists. It is particularly well-suited for use in individual student projects or as a simulation tool in one- or two-semester neural network-related courses at universities.

AI based optimal analysis of electro-osmotic peristaltic motion of non-Newtonian fluid with chemical reaction using artificial neural networks and response surface methodology

PurposeThe present work is devoted to investigating the sensitivity analysis of the electroosmotic peristaltic motion of non-Newtonian Casson fluid with the effect of the chemical reaction and magnetohydrodynamics through the porous medium. The main focus is on flow efficiency quantities such as pressure rise per wavelength, frictional forces on the upper wall and frictional forces on the lower wall. This initiative is to bridge the existing gap in the available literature.Design/methodology/approachThe governing equations of the problem are mathematically formulated and subsequently simplified for sensitivity analysis under the assumptions of a long wavelength and a small Reynolds number. The simplified equations take the form of coupled nonlinear differential equations, which are solved using the built-in Matlab routine bvp4c. The response surface methodology and artificial neural networks are used to develop the empirical model for pressure rise per wavelength, frictional forces on the upper wall and frictional forces on the lower wall.FindingsThe empirical model demonstrates an excellent fit with a coefficient of determination reaching 100% for responses, frictional forces on the upper wall and frictional forces on the lower wall and 99.99% for response, for pressure rise per wavelength. It is revealed through the sensitivity analysis that pressure rise per wavelength, frictional forces on the upper wall and frictional forces on the lower wall are most sensitive to the permeability parameter at all levels.Originality/valueThe objective of this study is to use artificial neural networks simulation and analyze the sensitivity of electroosmotic peristaltic motion of non-Newtonian fluid with the effect of chemical reaction.

Fully solution-driven charge trapping synaptic transistor with low energy consumption for neuromorphic computing

Brain-inspired neuromorphic computing has garnered significant attention for going beyond the constraint of von Neumann architecture. To emulate the human brain functions, various artificial synaptic devices have been proposed. Due to the high reliability and the CMOS compatibility, the synaptic transistors based on charge trapping (CT) mechanism have been considered to be one of the most promising candidates. However, most of the synaptic transistors based on CT mechanism were fabricated by costly vacuum-based techniques. In this report, based on a fully solution-driven strategy, the InZnO synaptic transistors, with Nd2O3 as the CT layer and ZrO2 as the dielectric layer, were integrated. The typical synaptic behaviors, including excitatory postsynaptic current, inhibitory postsynaptic current, memory enhancement, potentiation, and depression characteristics, were simulated by modulating presynaptic spikes. It is confirmed that the fabricated synaptic transistor shows low channel conductance and low energy consumption of 0.13 pJ per synaptic event. A recognition accuracy of 93.0% was achieved for the MNIST handwritten digital image dataset by an artificial neural network simulation. This study demonstrates the feasibility of solution-processed synaptic transistors, which exhibit significant potential for the neuromorphic applications.

Artificial Neural Network and Numerical Simulation for Magneto Hydrodynamics Hybrid Nanofluid Flow Towards a Stretching Cylinder

The present study focuses on influence of boundary layer flow of heat transfer analysis on hybrid nanofluid through a stretching cylinder. Moreover, the impressions of porous medium, heat generation/absorption are discussed. Further, the stimulus of non-linear thermal radiation and MHD are part of this investigation. For motivation, the Artificial Neural Networks also deliberated. Arrogate similarity variables are employed to transform the governing modelled PDEs into a couple of highly nonlinear ODEs. A numerical approach based on the shooting approach with MATLAB bvp5c built in function is employed for solution of the set of resulting ODEs and acquired outcomes are compared with existing literature, obtained results an exceptional agreement. Artificial nerve cells or node locations form a network of nerves, a contemporary name for a chain composed of human brain cells. The impact of physical parameters like Magnetic Field, Eckert number, Porosity, Prandtl number, Heat generation, thermal radiation, parameters on dimensionless velocity and energy fields are discussed through graphs and tables. The velocity profile decreased by about 42% when the magnetic field parameter values increases from 0.5 to 1.5. On the other hand increased by 78% on energy profile. The energy profile improved by about 97% when the Eckert number values increases from 1.0 < Ec < 3.0. The current model may be applicable in real life practical implications of employing Water–Cu/Ag/Al2O3 nanofluids on cylinders encompass enhanced heat transfer efficiency, and extended component lifespan, energy savings, and environmental benefits. This kind of theoretical analysis may be used in daily life applications, such as industrial and biomedical industries.

A boundary element method of bidomain modeling for predicting cellular responses to electromagnetic fields

Objective. Commonly used cable equation approaches for simulating the effects of electromagnetic fields on excitable cells make several simplifying assumptions that could limit their predictive power. Bidomain or ‘whole’ finite element methods have been developed to fully couple cells and electric fields for more realistic neuron modeling. Here, we introduce a novel bidomain integral equation designed for determining the full electromagnetic coupling between stimulation devices and the intracellular, membrane, and extracellular regions of neurons. Approach. Our proposed boundary element formulation offers a solution to an integral equation that connects the device, tissue inhomogeneity, and cell membrane-induced E-fields. We solve this integral equation using first-order nodal elements and an unconditionally stable Crank–Nicholson time-stepping scheme. To validate and demonstrate our approach, we simulated cylindrical Hodgkin–Huxley axons and spherical cells in multiple brain stimulation scenarios. Main Results. Comparison studies show that a boundary element approach produces accurate results for both electric and magnetic stimulation. Unlike bidomain finite element methods, the bidomain boundary element method does not require volume meshes containing features at multiple scales. As a result, modeling cells, or tightly packed populations of cells, with microscale features embedded in a macroscale head model, is simplified, and the relative placement of devices and cells can be varied without the need to generate a new mesh. Significance. Device-induced electromagnetic fields are commonly used to modulate brain activity for research and therapeutic applications. Bidomain solvers allow for the full incorporation of realistic cell geometries, device E-fields, and neuron populations. Thus, multi-cell studies of advanced neuronal mechanisms would greatly benefit from the development of fast-bidomain solvers to ensure scalability and the practical execution of neural network simulations with realistic neuron morphologies.

Graphene-Based Artificial Dendrites for Bio-Inspired Learning in Spiking Neuromorphic Systems.

Analog neuromorphic computing systems emulate the parallelism and connectivity of the human brain, promising greater expressivity and energy efficiency compared to those of digital systems. Though many devices have emerged as candidates for artificial neurons and artificial synapses, there have been few device candidates for artificial dendrites. In this work, we report on biocompatible graphene-based artificial dendrites (GrADs) that can implement dendritic processing. By using a dual side-gate configuration, current applied through a Nafion membrane can be used to control device conductance across a trilayer graphene channel, showing spatiotemporal responses of leaky recurrent, alpha, and Gaussian dendritic potentials. The devices can be variably connected to enable higher-order neuronal responses, and we show through data-driven spiking neural network simulations that spiking activity is reduced by ≤15% without accuracy loss while low-frequency operation is stabilized. This positions the GrADs as strong candidates for energy efficient bio-interfaced spiking neural networks.

One-Shot Remote Integration of Macromolecular Synaptic Elements on a Chip for Ultrathin Flexible Neural Network System.

The field of biomimetic electronics that mimicsynapticfunctionshas expanded significantly to overcome the limitationsofthevon Neumann bottleneck. However, the scaling down of the technology has led to an increasingly intricate manufacturing process. To address the issue, this work presents a one-shot integrable electropolymerization (OSIEP) method with remote controllability for the deposition of synaptic elements on a chip by exploiting bipolar electrochemistry. Condensing synthesis, deposition, and patterning into a single fabrication step is achieved by combining alternating-current voltage superimposed on direct-current voltage-bipolar electropolymerization and a specially designed dual source/drain bipolar electrodes. As a result, uniform 6 × 5 arrays of poly(3,4-ethylenedioxythiophene) channels are successfullyfabricatedon flexible ultrathin parylene substrates in one-shot process. The channels exhibited highly uniform characteristics and are directly used as electrochemical synaptic transistor with synaptic plasticity over 100 s. The synaptic transistors have demonstrated promising performance in an artificial neural network (NN) simulation, achieving a high recognition accuracy of 95.20%. Additionally, the array of synaptic transistor is easily reconfigured to a multi-gate synaptic circuit to implement the principles of operant conditioning. These results provide a compelling fabrication strategy for realizing cost-effective and disposable NN systems with high integration density.

Simulation and modeling of deep adversarial probabilistic neural network-based intrusion prevention system in cloud computing for smart grid cyber physical systems

Smart Grid Cyber Physical Systems (SG-CPS) have a substantial impact on power grid infrastructure upgrading. Nonetheless, due to the sophisticated nature of the infrastructure and the critical demand for resilient intrusion prevention systems, the task of protecting its security against data integrity attacks is a significant challenge. The simulation and assessment of security performance in SG-CPS present substantial hurdles in real-world power grid systems, owing mostly to experimental constraints. This necessitates the development of novel ways to improve distribution chain security. This research introduces a novel approach, employing a Deep Adversarial Probabilistic Neural Network (DAPNN)-based Intrusion prevention system in a cloud environment. Combining Bayesian Probabilistic Neural Networks (BEPNNs) with adversarial training and rule-based decision-making enhances precision and resilience. The major goal of this research is to detect and counteract false data injection (FDI) attacks that have the potential to compromise the integrity of power grid data. This paper proposes a novel methodology for intrusion detection in SG-CPS that combines BEPNNs with adversarial training. The addition of rule-based decision-making improves the system’s precision and resilience. The IEEE 24-bus system provides the foundation for providing data points relevant to normal operating conditions, contingency scenarios, and intentional attacks. The training procedure includes the use of a BEPNN for feature extraction, as well as the use of adversarial training approaches. The intrusion detection system has decision-making logic based on rules. The cloud infrastructure solution used in the study is Microsoft Azure. The results show that the DAPNN-based Intrusion Prevention System is effective in detecting and mitigating FDI attacks in SG-CPS. The system outperforms in terms of accuracy, precision, recall and F-measure, hence improving the security of the power grid infrastructure.

A comprehensive exploration of approximate DNN models with a novel floating-point simulation framework

This paper introduces TorchAxf11https://github.com/knuscent/TorchAxf., a framework for fast simulation of diverse approximate deep neural network (DNN) models, including spiking neural networks (SNNs). The proposed framework utilizes various approximate adders and multipliers, supports industrial standard reduced precision floating-point formats, such as bfloat16, and accommodates user-customized precision representations. Leveraging GPU acceleration on the PyTorch framework, TorchAxf accelerates approximate DNN training and inference. In addition, it allows seamless integration of arbitrary approximate arithmetic algorithms with C/C++ behavioral models to emulate approximate DNN hardware accelerators.We utilize the proposed TorchAxf framework to assess twelve popular DNN models under approximate multiply-and-accumulate (MAC) operations. Through comprehensive experiments, we determine the suitable degree of floating-point arithmetic approximation for these DNN models without significant accuracy loss and offer the optimal reduced precision formats for each DNN model. Additionally, we demonstrate that approximate-aware re-training can rectify errors and enhance pre-trained DNN models under reduced precision formats. Furthermore, TorchAxf, operating on GPU, remarkably reduces simulation time for complex DNN models using approximate arithmetic by up to 131.38× compared to the baseline optimized CPU implementation. Finally, we compare the proposed framework with state-of-the-art frameworks to highlight its superiority.

Innate Orientating Behavior of a Multi-Legged Robot Driven by the Neural Circuits of C. elegans.

The objective of this research is to achieve biologically autonomous control by utilizing a whole-brain network model, drawing inspiration from biological neural networks to enhance the development of bionic intelligence. Here, we constructed a whole-brain neural network model of Caenorhabditis elegans (C. elegans), which characterizes the electrochemical processes at the level of the cellular synapses. The neural network simulation integrates computational programming and the visualization of the neurons and synapse connections of C. elegans, containing the specific controllable circuits and their dynamic characteristics. To illustrate the biological neural network (BNN)'s particular intelligent control capability, we introduced an innovative methodology for applying the BNN model to a 12-legged robot's movement control. Two methods were designed, one involving orientation control and the other involving locomotion generation, to demonstrate the intelligent control performance of the BNN. Both the simulation and experimental results indicate that the robot exhibits more autonomy and a more intelligent movement performance under BNN control. The systematic approach of employing the whole-brain BNN for robot control provides biomimetic research with a framework that has been substantiated by innovative methodologies and validated through the observed positive outcomes. This method is established as follows: (1) two integrated dynamic models of the C. elegans' whole-brain network and the robot moving dynamics are built, and all of the controllable circuits are discovered and verified; (2) real-time communication is achieved between the BNN model and the robot's dynamical model, both in the simulation and the experiments, including applicable encoding and decoding algorithms, facilitating their collaborative operation; (3) the designed mechanisms using the BNN model to control the robot are shown to be effective through numerical and experimental tests, focusing on 'foraging' behavior control and locomotion control.