Computational Performance Research Articles

"Just Like, Risking Your Life Here": Participatory Design of User Interactions with Risk Detection AI to Prevent Online-to-Offline Harm Through Dating Apps

Social computing platforms facilitate interpersonal harms that manifest across online and physical realms such as sexual violence between online daters and sexual grooming through social media. Risk detection AI has emerged as an approach to preventing such harms, however a myopic focus on computational performance has been criticized in HCI literature for failing to consider how users should interact with risk detection AI to stay safe. In this paper we report an interview study with woman-identifying online daters (n=20) about how they envision interacting with risk detection AI and how risk detection models can be designed pursuant to such interactions. In accordance with this goal, we engaged women in risk detection model building exercises to build their own risk detection models. Findings show that women anticipate interacting with risk detection AI to augment - not replace - their personal risk assessment strategies. They likewise designed risk detection models to amplify their subjective and admittedly biased indicators of risk. Design implications involve the notion of personalizable risk detection models, but also ethical concerns around perpetuating problematic stereotypes associated with risk.

SMS-Net: Bridging the Gap Between High Accuracy and Low Computational Cost in Pose Estimation

Human pose estimation identifies and classifies key joints of the human body in images or videos. Existing pose estimation methods can precisely capture human movements in real time but require significant computational time and resources, which restricts their usage in specific conditions. Thus, we propose a lightweight pose estimation model—SMS-Net—based on the sequentially stacked structure of the hourglass network. The proposed model uses various lightweight techniques to enable high-speed pose estimation while requiring minimal storage space and computation. Specifically, a shuffle-gated block was introduced to reduce the computational load and number of parameters during the feature extraction process of the encoder composing each hourglass network. A multi-dilation block was used in the decoder to secure the receptive fields of various scales without increasing the computational load. The performance of the proposed model was assessed using the MPII and Common Objects in Context (COCO) datasets used for pose estimation and certain performance metrics and compared with state-of-the-art lightweight pose estimation models. Furthermore, an ablation study was performed to assess the impact of each module on network performance and efficiency. The results demonstrate that the proposed model achieved an improved balance between computational efficiency and performance compared to existing models in human pose estimation. Overall, the study findings can provide a basis for applications in computer vision technology.

Analysing the radiation reliability, performance and energy consumption of low-power SoC through heterogeneous parallelism

This study focuses on the low-power Tegra X1 System-on-Chip (SoC) from the Jetson Nano Developer Kit, which is increasingly used in various environments and tasks. As these SoCs grow in prevalence, it becomes crucial to analyse their computational performance, energy consumption, and reliability, especially for safety-critical applications. A key factor examined in this paper is the SoC’s neutron radiation tolerance. This is explored by subjecting a parallel version of matrix multiplication, which has been offloaded to various hardware components via OpenMP, to neutron irradiation. Through this approach, this researcher establishes a correlation between the SoC’s reliability and its computational and energy performance. The analysis enables the identification of an optimal workload distribution strategy, considering factors such as execution time, energy efficiency, and system reliability. Experimental results reveal that, while the GPU executes matrix multiplication tasks more rapidly and efficiently than the CPU, using both components only marginally reduces execution time. Interestingly, GPU usage significantly increases the SoC’s critical section, leading to an escalated error rate for both Detected Unrecoverable Errors (DUE) and Silent Data Corruptions (SDC), with the CPU showing a higher average number of affected elements per SDC.

Approximate single precision floating point adder for low power applications

With an increasing demand for power-hungry data-intensive computing, design methodologies with low power consumption are increasingly gaining prominence in the industry. Most of the systems operate on critical and noncritical data both. An attempt to generate a precision result results in excessive power consumption and results in a slower system. For noncritical data, approximate computing circuits significantly reduce the circuit complexity and hence power consumption. In this paper, a novel approximate single precision floating point adder is proposed with an approximate mantissa adder. The mantissa adder is designed with three 8-bit full adder blocks. In this paper, a detailed mathematical background, and proposed design approach in terms of the circuit configuration and truth tables are discussed. Additionally, a concept of switching between exact computing and approximate computing is analysed considering an approximate carry look-ahead adder. The delay and power consumption for the exact operating mode and approximate operation mode considering varied window sizes is observed. Performance of the approximate computation is compared against exact computation and varied approximate computing approaches.

Topology optimization designed twisted conformal cooling channel for additive-manufactured hot-stamping tool

Hot stamping is a manufacturing process that can be used to produce complex-shaped parts with high mechanical properties. In this process, the part quality and the production rate are strongly influenced by the efficiency of the cooling system within the tool. Improving the efficiency of the cooling system is therefore a key factor in reducing production time and costs. Traditional cooling systems for hot stamping tools often rely on simple drilled channels, limiting their potential for optimization. But recent improvements in additive manufacturing of steel have made it possible to design and produce cooling channels of more complex shapes and therefore allow the use of innovative design approaches such as topology optimization. This paper proposes a new rational method for optimum design of cooling system for additively manufactured hot stamping tools. The proposed new methodology is based on fluid-thermal topology optimization (TO). A low-fidelity model using the Stokes-Darcy equations is used to simulate fluid flow in the cooling channels. The objective function and constraints are designed to minimize the maximum temperature of the tool surface (cooling efficiency) with limited pressure drop of the cooling system. Advanced computations techniques including parallel computation, stabilization techniques, adaptive mesh and iterative solvers are used to ensure the computation performance. The procedure developed is first tested on a simple example to check the consistency of the results obtained. Then, to demonstrate the contribution of our work to the field under study, the procedure developed is used to optimize the design of the cooling system for an industrial hot stamping punch to be additively manufactured. The obtained optimum design is compared with standard drilled channels design and the deign with shell and core technology (with insert). These comparisons clearly shows that the optimal design combined with additive manufacturing can reduce quenching time by 37 % compared to conventional drilled straight channels, with very similar pressure drop and very similar load-bearing capacity. This research offers significant potential for improving efficiency and reducing costs in hot stamping and other manufacturing processes.

Deep Autoencoder for Real-Time Single-Channel EEG Cleaning and Its Smartphone Implementation Using TensorFlow Lite With Hardware/Software Acceleration.

To remove signal contamination in electroencephalogram (EEG) traces coming from ocular, motion, and muscular artifacts which degrade signal quality. To do this in real-time, with low computational overhead, on a mobile platform in a channel count independent manner to enable portable Brain-Computer Interface (BCI) applications. We propose a Deep AutoEncoder (DAE) neural network for single-channel EEG artifact removal, and implement it on a smartphone via TensorFlow Lite. Delegate based acceleration is employed to allow real-time, low computational resource operation. Artifact removal performance is quantified by comparing corrupted and ground-truth clean EEG data from public datasets for a range of artifact types. The on-phone computational resources required are also measured when processing pre-saved data. DAE cleaned EEG shows high correlations with ground-truth clean EEG, with average Correlation Coefficients of 0.96, 0.85, 0.70 and 0.79 for clean EEG reconstruction, and EOG, motion, and EMG artifact removal respectively. On-smartphone tests show the model processes a 4 s EEG window within 5 ms, substantially outperforming a comparison FastICA artifact removal algorithm. The proposed DAE model shows effectiveness in single-channel EEG artifact removal. This is the first demonstration of a low-computational resource deep learning model for mobile EEG in smartphones with hardware/software acceleration. This work enables portable BCIs which require low latency real-time artifact removal, and potentially operation with a small number of EEG channels for wearability. It makes use of the artificial intelligence accelerators found in modern smartphones to improve computational performance compared to previous artifact removal approaches.

Efficient Stochastic Template Bank Using Inner Product Inequalities

Gravitational wave searches are crucial for studying compact sources such as neutron stars and black holes. Many sensitive modeled searches use matched filtering to compare gravitational strain data to a set of waveform models known as template banks. We introduce a new stochastic placement method for constructing template banks, offering efficiency and flexibility to handle arbitrary parameter spaces, including orbital eccentricity, tidal deformability, and other extrinsic parameters. This method can be computationally limited by the ability to compare proposal templates with the accepted templates in the bank. To alleviate this computational load, we introduce the use of inner product inequalities to reduce the number of required comparisons. We also introduce a novel application of Gaussian Kernel Density Estimation to enhance waveform coverage in sparser regions. Our approach has been employed to search for eccentric binary neutron stars, low-mass neutron stars, primordial black holes, and supermassive black hole binaries. We demonstrate that our method produces self-consistent banks that recover the required minimum fraction of signals. For common parameter spaces, our method shows comparable computational performance and similar template bank sizes to geometric placement methods and stochastic methods, while easily extending to higher-dimensional problems. The time to run a search exceeds the time to generate the bank by a factor of O(105) for dedicated template banks, such as geometric, mass-only stochastic, and aligned spin cases, O(104) for eccentric and O(103) for the tidal deformable bank. With the advent of efficient template bank generation, the primary area for improvement is developing more efficient search methodologies.

Executive function is associated with the development of math performance in children born at

To describe associations between executive function (EF) domains (attentional control, information processing, cognitive flexibility, and goal setting) and concurrent math computation performance at age 7 and 13 years in children born <30 weeks' gestation or weighing <1,250 g, and second, to examine the impact of 7-year EF on math performance at 13 years. In a prospective, longitudinal cohort of children born <30 weeks' gestation or with a birthweight <1,250 g, assessment of EF and math performance was undertaken at 7 (n = 187) and 13 years (n = 174). Linear regression models were used to describe associations between EF domains with math performance at both time points, as well as to examine the impact of EF at 7 years on math performance at 13 years. At 7 and 13 years, all EF domains were positively and strongly associated with concurrent math performance (β = 11.35, 95% CI [9.28, 13.41] to β = 13.79, 95% CI [11.59, 15.98]). All EF domains at age 7 years were positively associated with math performance at 13 years, with the strongest associations observed for cognitive flexibility (β = 10.79 [8.64, 12.94]) and goal setting (β = 10.37 [8.08, 12.67]). This study provides evidence that EF is strongly associated with math performance in children born <30 weeks' gestation or with a birthweight <1,250 g and highlights the importance of early cognitive flexibility and goal setting performance for future math performance. (PsycInfo Database Record (c) 2024 APA, all rights reserved).

Leveraging multiplexed metasurfaces for multi-task learning with all-optical diffractive processors

Abstract Diffractive Neural Networks (DNNs) leverage the power of light to enhance computational performance in machine learning, offering a pathway to high-speed, low-energy, and large-scale neural information processing. However, most existing DNN architectures are optimized for single tasks and thus lack the flexibility required for the simultaneous execution of multiple tasks within a unified artificial intelligence platform. In this work, we utilize the polarization and wavelength degrees of freedom of light to achieve optical multi-task identification using the MNIST, FMNIST, and KMNIST datasets. Employing bilayer cascaded metasurfaces, we construct dual-channel DNNs capable of simultaneously classifying two tasks, using polarization and wavelength multiplexing schemes through a meta-atom library. Numerical evaluations demonstrate performance accuracies comparable to those of individually trained single-channel, single-task DNNs. Extending this approach to three-task parallel recognition reveals an expected performance decline yet maintains satisfactory classification accuracies of greater than 80 % for all tasks. We further introduce a novel end-to-end joint optimization framework to redesign the three-task classifier, demonstrating substantial improvements over the meta-atom library design and offering the potential for future multi-channel DNN designs. Our study could pave the way for the development of ultrathin, high-speed, and high-throughput optical neural computing systems.

Spectral cluster supertree: fast and statistically robust merging of rooted phylogenetic trees

The algorithms for phylogenetic reconstruction are central to computational molecular evolution. The relentless pace of data acquisition has exposed their poor scalability and the conclusion that the conventional application of these methods is impractical and not justifiable from an energy usage perspective. Furthermore, the drive to improve the statistical performance of phylogenetic methods produces increasingly parameter-rich models of sequence evolution, which worsens the computational performance. Established theoretical and algorithmic results identify supertree methods as critical to divide-and-conquer strategies for improving scalability of phylogenetic reconstruction. Of particular importance is the ability to explicitly accommodate rooted topologies. These can arise from the more biologically plausible non-stationary models of sequence evolution. We make a contribution to addressing this challenge with Spectral Cluster Supertree, a novel supertree method for merging a set of overlapping rooted phylogenetic trees. It offers significant improvements over Min-Cut supertree and previous state-of-the-art methods in terms of both time complexity and overall topological accuracy, particularly for problems of large size. We perform comparisons against Min-Cut supertree and Bad Clade Deletion. Leveraging two tree topology distance metrics, we demonstrate that while Bad Clade Deletion generates more correct clades in its resulting supertree, Spectral Cluster Supertree’s generated tree is generally more topologically close to the true model tree. Over large datasets containing 10,000 taxa and ∼500 source trees, where Bad Clade Deletion usually takes ∼2 h to run, our method generates a supertree in on average 20 s. Spectral Cluster Supertree is released under an open source license and is available on the python package index as sc-supertree.

On Interdicting Dense Clusters in a Network

Given a vertex-weighted undirected graph with blocking costs of its vertices and edges, we seek a minimum cost subset of vertices and edges to block such that the weight of any γ-quasi-clique in the interdicted graph is at most some predefined threshold parameter. The value of [Formula: see text] specifies the edge density of cohesive vertex groups of interest in the network. The considered weighted γ-quasi-clique interdiction problem can be viewed as a natural generalization of several variations of the clique blocker problem previously studied in the literature. From the application perspective, this setting is primarily motivated by the problem of disrupting adversarial (“dark”) networks (e.g., social or communication networks), where γ-quasi-cliques represent “tightly knit” groups of adversaries that we aim to dismantle. We first address the theoretical computational complexity of the problem. We then exploit some basic characterization of its feasible solutions to derive a linear integer programming (IP) formulation. This linear IP model can be solved using a lazy-fashioned branch-and-cut scheme. We also propose a combinatorial branch-and-bound algorithm for solving this problem. The computational performance of the developed exact solution schemes is studied using a test bed of randomly generated and real-life networks. Finally, some interesting insights and observations are also provided using a well-known example of a terrorist network. History: Accepted by Russel Bent, Area Editor for Network Optimization: Algorithms &amp; Applications. Funding: The work of S. Butenko was partially supported by the Air Force Office of Scientific Research under Award FA9550-23-1-0300. The work of O. A. Prokopyev was partially supported by the Office of Naval Research under Award ONR N00014-22-1-2678. Supplemental Material: The software that supports the findings of this study is available within the paper and its Supplemental Information ( https://pubsonline.informs.org/doi/suppl/10.1287/ijoc.2023.0027 ) as well as from the IJOC GitHub software repository ( https://github.com/INFORMSJoC/2023.0027 ). The complete IJOC Software and Data Repository is available at https://informsjoc.github.io/ . The online appendix is available at https://doi.org/10.1287/ijoc.2023.0027 .

Self-Powered Dye-Sensitized Solar-Cell-Based Synaptic Devices for Multi-Scale Time-Series Data Processing in Physical Reservoir Computing.

Physical reservoir computing (PRC) using synaptic devices has attracted attention as a promising edge artificial intelligence device. To handle time-series data on various time scales, it is necessary to fabricate devices with the desired time scale. In this study, we fabricated a dye-sensitized solar-cell-based synaptic device with controllable time constants by changing the light intensity. This device showed synaptic features, such as paired-pulse facilitation and paired-pulse depression, in response to light intensity. Moreover, we found that the high computational performance of the time-series data processing task was achieved by changing the light intensity, even when the input pulse width was varied. In addition, the fabricated device can be used for motion recognition tasks. This study paves the way for realizing multiple time-scale PRC.

Upper limb exoskeleton rehabilitation robot inverse kinematics modeling and solution method based on multi-objective optimization

The inverse kinematics problem of exoskeleton rehabilitation robots is challenging due to the lack of a standard analytical model, resulting in a complex and varied solution process. This complexity is especially pronounced in redundant upper limb exoskeleton robots, where inefficient solutions hinder the robot’s ability to adapt to the kinematic shape of the upper limb. This paper proposes a modeling and solution method based on multi-objective optimization to address the inverse kinematics of upper limb exoskeleton robots. We analyzed and validated this method using a redundant upper limb exoskeleton rehabilitation robot system developed by ourselves. First, we established a multi-objective inverse kinematics solution model by defining the end-position function, joint motion comfort function, system energy consumption function, motion safety, and human-like constraints. Then, the solution was designed based on the Improved Equilibrium Optimization (IEO) algorithm and validated its computational performance in terms of optimization ability, accuracy, and robustness. Finally, we experimentally tested the inverse kinematics solution model with the IEO algorithm on discrete objectives and continuous training trajectories using a redundant upper limb exoskeleton rehabilitation robot system. The results show that by incorporating the joint comfort function, system energy consumption function, and human-like constraints into the inverse kinematics model, we can not only quickly solve the inverse kinematics of redundant upper limb exoskeleton robots but also significantly improve the robot’s motion shape. Furthermore, it has better solution accuracy and stronger robustness than other algorithms when solving this inverse kinematics model based on the IEO algorithm.

MUCM-Net: a Mamba powered UCM-Net for skin lesion segmentation

Aim: Skin lesion segmentation is critical for early skin cancer detection. Challenges in automatic segmentation from dermoscopic images include variations in color, texture, and artifacts of indistinct lesion boundaries. This study aims to develop and evaluate MUCM-Net, a lightweight and efficient model for skin lesion segmentation, leveraging Mamba state-space models integrated with UCM-Net architecture optimized for mobile deployment and early skin cancer detection. Methods: MUCM-Net combines Convolutional Neural Networks (CNNs), multi-layer perceptions (MLPs), and Mamba elements into a hybrid feature learning module. Results: The model was trained and tested on the International Skin Imaging Collaboration (ISIC) 2017 and ISIC2018 datasets, consisting of 2,000 and 2,594 dermoscopic images, respectively. Critical metrics for evaluation included Dice Similarity Coefficient (DSC), sensitivity (SE), specificity (SP), and accuracy (ACC). The model’s computational efficiency was also assessed by measuring Giga Floating-point Operations Per Second (GFLOPS) and the number of parameters. MUCM-Net demonstrated superior performance in skin lesion segmentation with an average DSC of 0.91 on the ISIC2017 dataset and 0.89 on the ISIC2018 dataset, outperforming existing models. It achieved high SE (0.93), SP (0.95), and ACC (0.92) with low computational demands (0.055–0.064 GFLOPS). Conclusions: The model’s innovative Mamba-UCM layer significantly enhanced feature learning while maintaining efficiency that is suitable for mobile devices. MUCM-Net establishes a new standard in lightweight skin lesion segmentation, balancing exceptional ACC with efficient computational performance. Its ability to perform well on mobile devices makes it a scalable tool for early skin cancer detection in resource-limited settings. The open-source availability of MUCM-Net supports further research and collaboration, promoting advances in mobile health diagnostics and the fight against skin cancer. MUCM-Net source code will be posted on https://github.com/chunyuyuan/MUCM-Net.

Uniformization and bounded Taylor series in Newton-Raphsonmethod improves computational performance for a multistate transition model estimation and inference.

Multistate transition models (MSTMs) are valuable tools depicting disease progression. However, due to the complexity of MSTMs, larger sample size and longer follow-up time in real-world data, the computation of statistical estimation and inference for MSTMs becomes challenging. A bounded Taylor series in Newton-Raphson procedure is proposed which leverages the uniformization technique to derive maximum likelihood estimates and corresponding covariance matrix. The proposed method, namely uniformization Taylor-bounded Newton-Raphson, is validated in three simulation studies, which demonstrate the accuracy in parameter estimation, the efficiency in computation time and robustness in terms of different situations. This method is also illustrated using a large electronic medical record data related to statin-induced side effects and discontinuation.

Non-Periodic Quantized Model Predictive Control Method for Underwater Dynamic Docking

This study proposed an event-triggered quantized model predictive control (ETQMPC) method for the dynamic docking of unmanned underwater vehicles (UUVs) and human-occupied vehicles (HOVs). The proposed strategy employed a non-periodic control approach that initiated the non-linear model predictive control (NMPC) optimization and state sampling based on tracking errors and deviations from the predicted optimal state, thereby enhancing computing performance and system efficiency without compromising the control quality. To further conserve communication resources and improve information transfer efficiency, a quantitative feedback mechanism was employed for sampling and state quantification. The simulation experiments were performed to verify the effectiveness of the method, demonstrating excellent docking trajectory tracking performance, robustness against bounded current interference, and significant reductions in computational and communication burdens. The experimental results demonstrated that the method outperformed in the docking trajectory tracking control performance significantly improved the computational and communication performance, and comprehensively improved the system efficiency.

Transforming Image Super-Resolution: A ConvFormer-based Efficient Approach.

Recent progress in single-image super-resolution (SISR) has achieved remarkable performance, yet the computational costs of these methods remain a challenge for deployment on resource-constrained devices. In particular, transformer-based methods, which leverage self-attention mechanisms, have led to significant breakthroughs but also introduce substantial computational costs. To tackle this issue, we introduce the Convolutional Transformer layer (ConvFormer) and propose a ConvFormer-based Super-Resolution network (CFSR), offering an effective and efficient solution for lightweight image super-resolution. The proposed method inherits the advantages of both convolution-based and transformer-based approaches. Specifically, CFSR utilizes large kernel convolutions as a feature mixer to replace the self-attention module, efficiently modeling long-range dependencies and extensive receptive fields with minimal computational overhead. Furthermore, we propose an edge-preserving feed-forward network (EFN) designed to achieve local feature aggregation while effectively preserving high-frequency information. Extensive experiments demonstrate that CFSR strikes an optimal balance between computational cost and performance compared to existing lightweight SR methods. When benchmarked against state-of-the-art methods such as ShuffleMixer, the proposed CFSR achieves a gain of 0.39 dB on the Urban100 dataset for the x2 super-resolution task while requiring 26% and 31% fewer parameters and FLOPs, respectively. The code and pre-trained models are available at https://github.com/Aitical/CFSR.

Adaptive isogeometric topology optimization of shell structures based on PHT-splines

This paper proposes an adaptive isogeometric topology optimization framework for shell structures by utilizing a continuous density field represented as Polynomial splines over Hierarchical T-meshes (PHT-splines). This framework ensures an exact representation of shell structures, eliminating the geometric inaccuracies commonly associated with topology optimization. In the meanwhile, the meshes used for design and analysis are refined adaptively and locally along the density boundary to achieve a smooth material layout with reduced degrees of freedom (DOF). The adaptive sensitivity filter is tailored to the characteristics of PHT-splines, where the filter radius is determined automatically and adaptively, without the need to specify parameters in advance. Numerical experiments conducted on various shell structures validate the efficacy of the proposed adaptive framework. In comparison with isogeometric topology optimization based on non-adaptive cases (i.e. B-splines), the proposed adaptive framework demonstrates a notable enhancement in computational efficiency, with a 50%−80% reduction in running time and a 30%−60% reduction in DOF. Additionally, we offer a detailed comparison between PHT-splines and other locally refinable splines in the context of shell topology optimization. Numerical experiments exhibit that the efficient and localized refinement capability of PHT-splines provides advantages in both computational efficiency and structural performance for topology optimization.

VAN-DAMME: GPU-accelerated and symmetry-assisted quantum optimal control of multi-qubit systems

We present an open-source software package, VAN-DAMME (Versatile Approaches to Numerically Design, Accelerate, and Manipulate Magnetic Excitations), for massively-parallelized quantum optimal control (QOC) calculations of multi-qubit systems. To enable large QOC calculations, the VAN-DAMME software package utilizes symmetry-based techniques with custom GPU-enhanced algorithms. This combined approach allows for the simultaneous computation of hundreds of matrix exponential propagators that efficiently leverage the intra-GPU parallelism found in high-performance GPUs. In addition, to maximize the computational efficiency of the VAN-DAMME code, we carried out several extensive tests on data layout, computational complexity, memory requirements, and performance. These extensive analyses allowed us to develop computationally efficient approaches for evaluating complex-valued matrix exponential propagators based on Padé approximants. To assess the computational performance of our GPU-accelerated VAN-DAMME code, we carried out QOC calculations of systems containing 10 - 15 qubits, which showed that our GPU implementation is 18.4× faster than the corresponding CPU implementation. Our GPU-accelerated enhancements allow efficient calculations of multi-qubit systems, which can be used for the efficient implementation of QOC applications across multiple domains. Program summaryProgram Title: VAN-DAMMECPC Library link to program files::https://doi.org/10.17632/zcgw2n5bjf.1Licensing provisions: GNU General Public License 3Programming language: C++ and CUDANature of problem: The VAN-DAMME software package utilizes GPU-accelerated routines and new algorithmic improvements to compute optimized time-dependent magnetic fields that can drive a system from a known initial qubit configuration to a specified target state with a large (≈1) transition probability.Solution method: Quantum control, GPU acceleration, analytic gradients, matrix exponential, and gradient ascent optimization.

Resource-optimized cnns for real-time rice disease detection with ARM cortex-M microprocessors

This study explores the application of Artificial Intelligence (AI), specifically Convolutional Neural Networks (CNNs), for detecting rice plant diseases using ARM Cortex-M microprocessors. Given the significant role of rice as a staple food, particularly in Malaysia where the rice self-sufficiency ratio dropped from 65.2% in 2021 to 62.6% in 2022, there is a pressing need for advanced disease detection methods to enhance agricultural productivity and sustainability. The research utilizes two extensive datasets for model training and validation: the first dataset includes 5932 images across four rice disease classes, and the second comprises 10,407 images across ten classes. These datasets facilitate comprehensive disease detection analysis, leveraging MobileNetV2 and FD-MobileNet models optimized for the ARM Cortex-M4 microprocessor. The performance of these models is rigorously evaluated in terms of accuracy and computational efficiency. MobileNetV2, for instance, demonstrates a high accuracy rate of 97.5%, significantly outperforming FD-MobileNet, especially in detecting complex disease patterns such as tungro with a 93% accuracy rate. Despite FD-MobileNet’s lower resource consumption, its accuracy is limited to 90% across varied testing conditions. Resource optimization strategies highlight that even slight adjustments, such as a 0.5% reduction in RAM usage and a 1.14% decrease in flash memory, can result in a notable 9% increase in validation accuracy. This underscores the critical balance between computational resource management and model performance, particularly in resource-constrained settings like those provided by microcontrollers. In summary, the deployment of CNNs on microcontrollers presents a viable solution for real-time, on-site plant disease detection, demonstrating potential improvements in detection accuracy and operational efficiency. This study advances the field of smart agriculture by integrating cutting-edge AI with practical agricultural needs, aiming to address the challenges of food security in vulnerable regions.