As physical AI systems such as service robots and autonomous monitoring agents are increasingly deployed in public spaces, ensuring the safety and well-being of humans they interact with remains a critical challenge. Existing approaches typically address verbal threat signals or visual threats in isolation, lacking the cross-modal reasoning necessary for complex real-world scenarios. In this ongoing research exploration, we present MindGuard, a multimodal agentic pipeline that unifies mental and physical threat detection through layered reasoning over text and vision signals. The system processes two parallel branches: a text branch combining keyword detection with a fine-tuned BERT classifier for mental health threat assessment, and a vision branch combining real-time object detection with a Vision-Language Model (VLM) for contextual scene understanding. Both signal streams are fused by an LLM-based agent that performs cross-modal analysis - recognizing, for example, that “I want to die - laughing” paired with a smiling person is a false positive, while “I’m fine” paired with a visible weapon requires immediate escalation. The fusion agent produces unified decisions across five severity levels with natural language responses, appropriate actions, and audit logs with camera snapshots for human review. An autonomous background monitor continuously scans the environment without requiring user input, enabling always-on surveillance essential for physical AI deployment. We evaluate MindGuard on 150 test scenarios spanning five threat categories, achieving an overall detection accuracy of 92.7% compared to 68.2% (text-only) and 67.9% (vision-only), while reducing false positive rates by through cross-modal validation. The entire system runs on local hardware using open-source models with no cloud dependency. We validate edge deployability by benchmarking the full pipeline on an NVIDIA Jetson Orin AGX, achieving 9.3-second end-to-end latency sufficient for continuous background monitoring. Logged interactions generate labeled datasets for classifier retraining, creating a self-improving safety system that bridges AI safety research and real-world physical AI deployment.

Event-based vision sensors such as the DVS346 capture asynchronous streams of pixel-level brightness changes with ultrahigh temporal resolution and sparse representation, which makes them ideal for dynamic perception tasks. However, their output format is not inherently compatible with spiking neural networks (SNNs), which require structured spike train input to utilize temporal coding. In this ongoing work, we present a robust conversion framework that algorithmically transforms raw event streams collected from a DVS346 camera mounted on a Broxter robotic platform into spike train representations suitable for SNN processing. Our method preserves critical temporal dynamics from the original event data while enabling compatibility with neuromorphic learning paradigms. To facilitate model interpretation and validation, we also introduce a visualization suite that reconstructs both event and generated spike datasets into frame-based videos for qualitative analysis. The resulting dataset bridges the gap between neuromorphic sensing and computation, promoting efficient learning and real-time inference in SNN models, and advancing the integration of event-based vision into embodied robotic systems.

The selection of optimal neural models in Spiking Neural Networks (SNNs) traditionally depends on a trial-and-error approach, which is both time-consuming and sometimes tends to suboptimal selection of the neural model. This research study explores the integration of Automated Machine Learning (AutoML) techniques into SNNs to simplify the process of designing an SNN model by automatically selecting the most optimal neural model. For example, in traditional neural networks, AutoML proves that it has the potential to significantly reduce the time and effort required to identify the best-performing architectures, by automating the selection and updating of loss functions and hyper-parametric values. To this end, we present a stand-alone framework for applying AutoML in SNNs, highlighting how this approach can improve accuracy, efficiency, and reliability in the selection process of optimal neural models, as these neural models are the core principle of designing SNN architecture. The proposed method not only accelerates the development of SNN models but also enhances their performance by systematically identifying optimal configurations that might be overlooked through manual methods. Therefore, to validate our approach, the proposed architecture was tested using different well-known benchmarks with different sets of neurons, ranging from 100 to 3000 neurons, and demonstrated state-of-the-art results in image classification tasks.

Automatic synthesis of approximate accelerators remains a challenging problem owing to extremely large design space and high verification costs for evaluating intermediate design candidates. Existing frameworks typically employ searchbased strategies to explore the design space. While some frameworks aggressively prune candidate configurations to reduce runtime, leading to prematurely eliminating high-quality solutions, others use a learning based approach for design space exploration with an increased runtime cost. In this work, we propose an advanced Design Space Exploration (DSE) approach that incorporates a refined, cost-aware selection and expansion strategy to enable a more balanced and effective traversal of the search space. Our method builds upon an existing learning-based state-of-the-art technique to improve the accuracy and quality of tradeoffs, especially for large-scale accelerators. Experimental results demonstrate that our proposed approach achieves comparable area and power savings for small to medium-scale benchmarks while significantly outperforming the state-of-the-art approaches on larger circuits with up to 9% additional area savings with almost no runtime overhead.

This article introduces the enhanced Runtime Analyzing and Visualization Simulator (RAVSim) v2.0, a graphical tool that not only supports SNN design and analysis but also facilitates a comprehensive comparative analysis of various SNN models. The new version of RAVSim introduces a groundbreaking feature enabling users to conduct in-depth comparisons of SNN models, enhancing understanding and aiding in model selection for specific applications. Furthermore, with the updated version of RAVSim, researchers, and developers can effortlessly generate trained model weights using a custom dataset, eliminating the need to investigate or write complicated backend code. This new feature facilitates the seamless integration of diverse datasets, streamlining the process for further analysis and exploration. Therefore, the developers can now focus on high-level tasks and gain a clear understanding of SNN without worrying about the technical complexities of weight generation. This advancement represents a significant step towards making SNNs more accessible and user-friendly, unlocking their full potential in artificial intelligence and computational neuroscience applications. Furthermore, RAVSim’s code has undergone extensive optimization and debugging, leading to a substantial ∼65% reduction in image classification simulation time compared to the previous RAVSim version. This improvement makes it easier and quicker to train models and generate weights.

The era of digitization and IoT devices is marked by the constant storage of massive amounts of data. The growing adoption of smart home environments, which use sensors and devices to monitor and control various aspects of daily life, underscores the need for effective privacy and security measures. HE is a technology that enables computations on encrypted data, preserving confidentiality. As a result, researchers have developed methodologies to protect user information, and HE is one of the technologies that make it possible to perform computations directly on encrypted data and produce results using this encrypted information. Thus, this research study compares the performance of three ML models, XGBoost, Random Forest, and Decision Classifier, on a real-world smart home dataset using both with and without FHE. Practical results demonstrate that the Decision Classifier showed remarkable results, maintaining high accuracy with FHE and even surpassing its plaintext performance, suggesting that encryption can enhance model accuracy under certain conditions. Additionally, Random Forest showed efficiency in terms of execution time and low prediction errors with FHE, making it a strong candidate for encrypted data processing in smart homes. These findings highlight the potential of FHE to set new privacy standards, advancing secure and privacy-preserving technologies in smart environments.

In this article, we propose a novel standalone hybrid Spiking-Convolutional Neural Network (SC-NN) model and test on using image inpainting tasks. Our approach uses the unique capabilities of SNNs, such as event-based computation and temporal processing, along with the strong representation learning abilities of CNNs, to generate high-quality inpainted images. The model is trained on a custom dataset specifically designed for image inpainting, where missing regions are created using masks. The hybrid model consists of SNNConv2d layers and traditional CNN layers. The SNNConv2d layers implement the leaky integrate-and-fire (LIF) neuron model, capturing spiking behavior, while the CNN layers capture spatial features. In this study, a mean squared error (MSE) loss function demonstrates the training process, where a training loss value of 0.015, indicates accurate performance on the training set and the model achieved a validation loss value as low as 0.0017 on the testing set. Furthermore, extensive experimental results demonstrate state-of-the-art performance, showcasing the potential of integrating temporal dynamics and feature extraction in a single network for image inpainting.

With the motivation and the difficulties that currently exist in comprehending and utilizing the promising features of SNNs, we proposed a novel run-time multi-core architecture-based simulator called "RAVSim" (Runtime Analysis and Visualization Simulator), a cutting-edge SNN simulator, developed using LabVIEW and it is publicly available on their website as an official module. RAVSim is a runtime virtual simulation environment tool that enables the user to interact with the model, observe its behavior of output concentration, and modify the set of parametric values at any time while the simulation is in execution. Recently some popular tools have been presented, but we believe that none of the tools allow users to interact with the model simulation in run time.

Spiking Neural Networks (SNNs) are a type of neural network designed to replicate biological neural networks more closely by using discrete spikes to transmit information. Unlike traditional etworks, SNNs incorporate time by relying on the precise timing of spikes for neuron-to-neuron communication. This reduces hardware omplexity, as it only requires one-bit logic, making SNNs ideal for hardware integration. This study assesses the performance of several SNN models for hardware implementation, focusing on resource utilization, speed, and power consumption. Verilog was used for the hardware design, and the simulations were run in Vivado. The emulation experiments were conducted on the Basys3 FPGA board to validate our findings. Our analysis indicates that simpler models like Leaky Integrate and Fire (LIF) and Non-linear Integrateand- Fire (NLIF) are highly efficient, with low resource and power requirements, making them suitable for resource-constrained environments. More complex models like Hodgkin-Huxley (HH) and Izhikevich (IZH) provide detailed neuronal dynamics but at a higher resource cost. Our implementations exhibit notable improvements across several metrics compared to previouswork. This analysis equips researchers with the necessary information to make informed decisions about which neuron model best meets their application needs, whether prioritizing speed, efficiency, or biological accuracy.

In the era of digitization and smart technologies, the generation of huge amounts of data has become ubiquitous. Extracting and mining valuable information from this data is both crucial and challenging. This paper proposes a hybrid approach that integrates fuzzy logic with deep learning techniques for anomaly detection and next-activity prediction in smart homes. The proposed methodology is particularly designed to support elderly residents by accurately detecting anomalous patterns and predicting daily activities, thereby enhancing the system's accuracy and reliability. The hybrid model uses fuzzy logic to define membership functions, rule-based decision making, and effectively handle abnormalities, while deep learning techniques are employed for predictive analysis. Experiments conducted on a real-world smart home data-set demonstrate that the proposed approach significantly outperforms traditional methods in terms of accuracy, correctness, and loss. The proposed research study makes a valuable contribution to the field of smart home technologies.

Spiking Neural Networks (SNNs), are inspired by the biological brain's complicated signaling mechanisms and possess unique characteristics that set them apart from traditional artificial neural networks. This research study explores the challenging domain of image classification, specifically utilizing the well-known MNIST dataset through the development and thorough evaluation of different neural models for edge computing. However, the primary contribution is the autonomous selection of the best-performing SNN model through various early stopping approaches and validation functions, allowing the models to autonomously adapt during training. In addition, this article presents the standalone AutoML-SNN model, which is the introduction of dynamic elements into selected SNN domains, enhancing their adaptability to complex patterns within the dataset. Furthermore, the early stopping methodologies are used to reduce overfitting hazards, and using the 3000-neuron set, the LIF appeared as the most proficient neural model.

This abstract introduces a Next-Gen runtime tool research direction for neural network programming, which was presented in the Programming 2024 event. These runtime revolutionary tools allow users to conduct neural model evaluations and understand neuron behavior in run-time, neural network design, and training processes. By eliminating the need for extensive backend coding, one of the previously introduced runtime simulators called RAVSim, simplified the integration of custom datasets and enabled developers to focus on high-level tasks, accelerating advancements in artificial intelligence (AI) and computational neuroscience. With its innovative features and user-friendly interface, RAVSim supports researchers and developers in using SNNs full potential in diverse applications.

Spiking Neural Networks (SNNs) have gained significant attention in the field of neuromorphic computing for their potential to mimic the brain's spiking neurons, allowing event-driven processing based on exact spike timing. In this paper, we introduce a novel architecture that uses the power of SNN in combination with transfer learning to achieve real-time human presence detection and analysis using event-based cameras and compare it with non-event-based cameras. This architecture, which is deployed on edge computing devices, controls a comprehensive pipeline of components, seamlessly integrating various strategies. It combines object detection, transfer learning with SNN, human recognition, localizing and tracking, feature extraction, multi-core architecture, and run-time analysis. The application is initiated by extensively detecting objects and monitoring environments for motion events. Thus, transfer learning adjusts pre-trained Convolutional Neural Network (CNN) weights to SNNs upon detection, enabling event-driven processing. The utilization of multi-core processing speeds up the analytical workload while maintaining real-time operations. The architecture also keeps a valuable spike train dataset, which records important information about recognized objects. This dataset is useful for applications such as behavioral analysis and real-time monitoring.

In recent years, as urban AI applications increasingly rely on sensitive data, ensuring the privacy and security of machine learning (ML) models has become essential. The proposed research study evaluates the performance and security trade-offs of seven encryption techniques applied to ML models used in urban AI settings. We compare encryption methods, including mixed homomorphic encryptions using Convolutional Neural Networks (CNNs) trained on the MNIST dataset, we analyze how these encryption methods affect model performance in terms of accuracy, error rate, and information leakage. The CNN models, after being trained with encrypted data, are deployed on embedded devices to evaluate real-time performance. We measure key metrics, including execution time, memory usage, and classification accuracy, to assess the feasibility of each encryption method in urban AI scenarios. Additionally, the impact of encryption on model interpretability and robustness is considered, particularly when used in urban applications like intelligent transportation systems, smart city sensors, and surveillance systems. By evaluating error rates, mutual information scores, and statistical properties such as mean and variance, this research aims to explore the practical trade-offs between security, privacy, and performance. Our findings highlight the importance of selecting appropriate encryption techniques for urban AI tasks to maintain both data privacy and model efficiency in real-world settings.

The advent of smart home technologies has opened new avenues for personalized healthcare, energy management, and enhanced convenience. However, variations in resident behavior, sensor types, and home layouts cause domain shifts, posing significant challenges to the deployment of machine learning models across diverse smart home environments. This paper describes a new method that uses domain adaptation layers in a transfer learning framework to change domain-specific features on the fly. This makes cross-data set generalization better. We compare the proposed method against Standard Transfer Learning, Domain-Adversarial Neural Networks (DANN), and Correlation Alignment (CORAL) using multiple smart home data sets. Our evaluation demonstrates that the proposed method significantly outperforms these existing approaches in terms of predictive performance, highlighting its effectiveness in addressing domain shifts in smart home environments.

This paper presents the results of our first assessment on the emulation of spiking neural networks (SNNs) on Field Programmable Gated Arrays (FPGAs). Three fundamental modules are designed to obtain a fully functional neural network. First, a module that provides the simulation of a single neuron’s properties, characteristics, and behavior using an accurate but, at the same time, a computationally efficient mathematical model is designed. Second, a communication or transmission system called Address Event Representation (AER) is developed to manage the neural network’s information flow between the neurons. Third, a training algorithm - Spike-Timing Dependent Plasticity (STDP), is designed to provide functionality to the neural network. This modular approach provides the necessary flexibility and scalability for simulating various SNNs and the different numbers of neurons. The modules have been implemented using multiple combinational blocks and flip-flops; The network description is performed using Very High-Speed Integrated Circuit Hardware Description Language (VHDL) in the Xilinx Vivado simulator. The SNN model for digit recognition is implemented and emulated on the Basys3 FPGA development board to demonstrate the accuracy of the model’s operation.

Graph Neural Networks (GNNs) are specialized neural networks that operate on graph-structured data, utilizing the connections between nodes to learn and process information. To achieve optimal performance, GNNs require the automatic selection of the best loss and optimization functions, which allows the model to adapt to the unique features of the dataset being used. This eliminates the need for manual selection, saving time and minimizing the requirement for domain-specific knowledge. The automatic selection of loss and optimization functions is a crucial factor in achieving state-of-the-art results when training GNNs. In this study, we trained Graph Convolutional Networks (GCNs) and Graph Attention Networks (GAT) models for node classification on three benchmark datasets. To automatically select the best loss and optimization functions, we utilized performance metrics. We implemented a learning rate scheduler to adjust the learning rate based on the model's performance, which led to improved results. We evaluated the model's performance using multiple metrics and reported the best loss function and performance metric, enabling users to compare its performance to other models. Our approach achieved state-of-the-art results, highlighting the importance of selecting the appropriate loss and optimizer functions. Additionally, we developed a real-time visualization of the GCN model during training, providing users with a detailed understanding of the model's behavior. Overall, this study provides a comprehensive understanding of GNNs and their application to graph-structured data, with a specific focus on real-time visualization of GNN behavior during training.

Spiking Neural Networks (SNNs) help achieve brain-like efficiency and functionality by building neurons and synapses that mimic the human brain's transmission of electrical signals. However, optimal SNN implementation requires a precise balance of parametric values. To design such ubiquitous neural networks, a graphical tool for visualizing, analyzing, and explaining the internal behavior of spikes is crucial. Although some popular SNN simulators are available, these tools do not allow users to interact with the neural network during simulation. To this end, we have introduced the first runtime interactive simulator, called Runtime Analyzing and Visualization Simulator (RAVSim),a developed to analyze and dynamically visualize the behavior of SNNs, allowing end-users to interact, observe output concentration reactions, and make changes directly during the simulation. In this paper, we present RAVSim with the current implementation of runtime interaction using the LIF neural model with different connectivity schemes, an image classification model using SNNs, and a dataset creation feature. Our main objective is to primarily investigate binary classification using SNNs with RGB images. We created a feed-forward network using the LIF neural model for an image classification algorithm and evaluated it by using RAVSim. The algorithm classifies faces with and without masks, achieving an accuracy of 91.8% using 1000 neurons in a hidden layer, 0.0758 MSE, and an execution time of ∼10[Formula: see text]min on the CPU. The experimental results show that using RAVSim not only increases network design speed but also accelerates user learning capability.

This article presents a comprehensive analysis of spiking neural networks (SNNs) and their mathematical models for simulating the behavior of neurons through the generation of spikes. The study explores various models, including LIF and NLIF, for constructing SNNs and investigates their potential applications in different domains. However, implementation poses several challenges, including identifying the most appropriate model for classification tasks that demand high accuracy and low-performance loss. To address this issue, this research study compares the performance, behavior, and spike generation of multiple SNN models using consistent inputs and neurons. The findings of the study provide valuable insights into the benefits and challenges of SNNs and their models, emphasizing the significance of comparing multiple models to identify the most effective one. Moreover, the study quantifies the number of spiking operations required by each model to process the same inputs and produce equivalent outputs, enabling a thorough assessment of computational efficiency. The findings provide valuable insights into the benefits and limitations of SNNs and their models. The research underscores the significance of comparing different models to make informed decisions in practical applications. Additionally, the results reveal essential variations in biological plausibility and computational efficiency among the models, further emphasizing the importance of selecting the most suitable model for a given task. Overall, this study contributes to a deeper understanding of SNNs and offers practical guidelines for using their potential in real-world scenarios.