Real-time classification or anomaly detection tasks on sensor data often represent a significant challenge, due to signal variability and noise levels corrupting the data in unpredictable ways. This issue is especially relevant in biomedical applications, where data acquisition is noticeably unbalanced, inter-person and inter-day signal variability is non-negligible, and data annotation is an extremely complex task to be completed by expert physicians. In this context, Spiking Neural Networks (SNNs) provide the interesting opportunity of mimicking the learning capabilities of the brain through different plasticity mechanisms, which allow an online refinement of the model’s parameters in an unsupervised fashion. This work aims at leveraging plasticity to adjust on the fly the classification performance of a lightweight SNN model, targeting biological signal monitoring in the wearable domain, by relying on a specialized low-power SNN accelerator enabling signal classification within a power envelope of 0.25 mW. The efficacy of the approach is evaluated on a seizure recognition problem based on the real-time processing of the electroencephalography (EEG) signal, where ensuring a good trade-off between anomaly detection and false-alarms minimization is paramount, and seizure occurrences vary significantly. The plasticity rule is executed on the host RISC-V processor integrated in the FPGA-based accelerator, enabling efficient and self-adaptive always-on monitoring with a negligible impact on the system’s power consumption.

Accurate and energy-efficient classification of cardiac arrhythmias is essential for real-time electrocardiogram (ECG) monitoring in wearable healthcare systems. This work introduces an end-to-end, spike-driven approach for ECG analysis, in which a spiking neural network (SNN) addresses arrhythmia detection on an event-based input. The signal encoding employs delta modulation on the raw ECG waveform as well as its first and second derivatives, capturing richer temporal and morphological features and enhancing classification performance compared to baseline approaches. Heartbeats are classified into five categories, as defined by the AAMI standard, achieving 98.4% accuracy on the MIT-BIH Arrhythmia Database. Unlike traditional methods, our approach removes the need for separate filtering, segmentation, or peak detection algorithms, relying instead on a unified, event-driven architecture. To support this enhanced processing methodology, we have implemented an optimized hardware architecture based on a low-power Lattice iCE40-UltraPlus Field Programmable Gate Array (FPGA). This design eliminates redundant computations by unifying peak detection and classification within the same processing pipeline, reducing power consumption while maintaining low inference times. Performance evaluations indicate an execution time of just 4.05 ms per classification, with energy usage optimized to 36.86 mu J, substantially outperforming existing FPGA-based solutions.

Accurate and energy-efficient recognition of swimming styles is essential for real-time training analysis in wearable systems. This work introduces an end-to-end, spike-driven approach for swimming style classification, in which a Spiking Neural Network (SNN) operates on event-based inertial inputs. To the best of our knowledge, this is the first application of neuromorphic computing to swimming style recognition. The signal encoding employs delta modulation on accelerometer, gyroscope, and magnetometer axes, converting continuous sensor streams into sparse binary spike trains that capture temporal dynamics of swimming movements without requiring explicit filtering or feature extraction. A compact fully connected SNN with leaky integrate-and-fire neurons classifies four swimming styles (freestyle, breaststroke, backstroke, and butterfly) using only 18,176 trainable parameters. The approach is validated on a publicly available benchmark of 40 swimmers recorded in realistic training conditions, achieving an F1 of 0.945 under Leave-One-Subject-Out cross-validation. To support deployment on resource-constrained wearable devices, we implement an optimized hardware architecture on a low-power Lattice iCE40-UltraPlus FPGA. The event-driven nature of the SNN enables efficient inference by processing only active spikes, reducing power consumption while maintaining low latency. Performance evaluations indicate an execution time of 7.8 ms per classification, with energy usage of 93.6 J, demonstrating the feasibility of on-device real-time swimming analysis.

Spiking Neural Networks (SNNs) exploit event-driven processing to offer high energy efficiency when deploying Artificial Intelligence (AI) on wearable edge devices. However, specialized hardware is needed to fully take advantage of this potential, which, despite recent advances, remains expensive and not widely accessible. To address this, open-source Electronic Design Automation (EDA) tools and Process Design Kits (PDKs) offer a path to democratize the development of neuromorphic hardware. In this work, we present SYNtzulA, a system-on-chip designed for SNN acceleration, developed using the open-source IHP-SG13G2 130 nm PDK and the OpenROAD toolchain. The chip integrates a RISC-V softcore and a dedicated SNN accelerator, occupying approximately 6.8 mm2 including I/O pads. It operates at up to 125 MHz, reaching a throughput of 2 Giga Synaptic Operations per second (GSOP/s) with an energy consumption of 36.5 pJ per synaptic operation. The accelerator can exploit the sparsity of spike-based computation by skipping unnecessary operations, resulting in total energy consumption in the order of a few hundred nanojoules per inference in different use cases involving biosignal analysis.

Wireless Capsule Endoscopy (WCE) offers an important diagnostic instrument for different gastrointestinal diseases. Enhancing the WCE device with real-time image processing capabilities allows to assist specialized physicians in the long and cumbersome process of inspecting the significant amount of data acquired during the examination procedure, providing the first detection of the signs of relevant diseases that require further attention. In this work, we evaluate different state-of-the-art Convolutional Neural Network models for real-time WCE image classification, focusing on lightweight topologies suitable for execution on low-power microcontroller platforms and integration on the WCE device. The selected WCE-SqueezeNet model achieves 98.5

Advancements in CMOS microelectrode array sensors have significantly improved sensing area and resolution, paving the way to accurate Brain-Machine Interfaces (BMIs). However, near-sensor neural decoding on implantable computing devices is still an open problem. A promising solution is provided by Spiking Neural Networks (SNNs), which leverage event sparsity to improve energy consumption. However, given the typical data rates involved, the workload related to I/O acquisition and spike encoding is dominant and limits the benefits achievable with event-based processing. In this work, we present two power-efficient implementations, on FPGA and ASIC, of a dedicated processor for the decoding of intracortical action potentials from primary motor cortex. The processor leverages lightweight sparse SNNs to achieve state-of-the-art accuracy. To limit the impact of I/O transfers on energy efficiency, we introduced a channel selection scheme that reduced bandwidth requirements by 3x and power consumption by 2.3x and 1.6x on the FPGA and ASIC, respectively, enabling inference at 0.446 mu J and 1.04 mu J, with no significant loss in accuracy. To promote broad adoption in a specialized, research-intensive domain, we have based our implementations on open-source EDA tools, low-cost hardware, and an open PDK.

The development of epilepsy monitoring solutions suitable for everyday use is a very challenging task, where different constraints should be combined, resulting from the required accuracy standards, the unobtrusiveness of the monitoring device, and the efficiency of real-time operation. Considering the time-varying nature of the electroencephalography signal (EEG), Spiking Neural Networks (SNNs) represent a promising solution to model the evolution of the brain state based on the history of the previously processed signal. This work proposes an extremely lightweight SNN-based seizure detection solution, utilizing a simple encoding scheme to ensure high levels of sparsity. Despite the reduced complexity, the model provides a detection performance comparable with the state-of-the-art SNN-based approaches on the evaluated data from the CHB-MIT dataset, reaching a 96% area under the curve (AUC) and allowing 99.3% average accuracy, with the detection of 100% of the examined seizure events and a false alarm rate of 0.3 false positives per hour. The suitability for real-time inference execution on wearable monitoring devices was assessed on SYNtzulu, demonstrating 0.5 μs inference time with 4.55 nJ energy consumption.

Spiking Neural Networks (SNNs) represent a promising class of neural networks that emulate the behavior of biological neurons, offering significant advantages in terms of energy efficiency and computational power. These networks achieve optimal performance on neuromorphic processors, however, such hardware remains constrained by prohibitive costs and limited accessibility. To address this gap, open source Process Design Kits (PDKs) and Electronic Design Automation (EDA) tools can be used to facilitate broader access to hardware development. In this work, we present SYNtzulA (SYNtzulu on ASIC), a system-on-chip that integrates a RISC-V softcore with a dedicated accelerator for SNNs. The chip layout was developed using the open-source IHP-SG13G2 PDK and the OpenROAD flow, reducing development costs and promoting wider accessibility to neuromorphic hardware solutions. SYNtzulA occupies an area of approximately 5.2 mm(2) and operates at a maximum frequency of 125 MHz. The system is capable of processing 109 synapses per second at its maximum frequency, while maintaining a power consumption during inference that scales linearly with the operating frequency, dissipating approximately 632 mu W/MHz.

The accurate modeling of hand movement based on the analysis of surface electromyographic (sEMG) signals offers exciting opportunities for the development of complex prosthetic devices and human-machine interfaces, moving from discrete gesture recognition, towards continuous movement tracking. In this study, we present two solutions for real-time sEMG processing, based on lightweight Spiking Neural Networks (SNNs) and efficiently implemented on a Lattice iCE40-UltraPlus FPGA, especially suitable for low-power applications. We first assess the performance in the discrete finger gesture recognition task, considering as a reference the NinaPro DB5 dataset, and demonstrating an accuracy of 83.17% in the classification of twelve different finger gestures. We also consider the more challenging problem of continuous finger force modeling, referencing the Hyser dataset for finger tracking during independent extension and contraction exercises. The assessment reveals a correlation of up to 0.875 with the ground-truth forces. Our systems take advantage of SNNs' inherent efficiency and, dissipating 11.31 mW in active mode, consume 44.6 mu J for a gesture recognition classification and 1.19 mu J for a force modeling inference. Considering dynamic power-consumption management and the introduction of idle periods, average power drops to 1.84 mW and 3.69 mW for these respective tasks.

Spiking Neural Networks (SNNs) are computational efficient algorithms that have proven to be very effective on edge applications. Recent efforts of the scientific community have investigated the integration of the attention layer into Spiking Neural Networks to close the accuracy gap with non-spiking approaches. As effective, hardware implementations of spiking transformers have remained limited to high-end FPGA platforms so far, with power requirements incompatible for low-power applications. This work introduces ESTU, a compact spiking transformer hardware architecture tailored for edge deployment. ESTU has been designed with resource minimization in mind in order to fit in power and resource constrained FPGAs, as the Lattice iCE40UP5K. We validated ESTU on available Spiking Transformer models at the State of the Art and on a sEMG classification application addressing NinaPro DB-5, demonstrating improved accuracy of over 2% compared to the results of a vanilla SNN, fully unlocking spiking transformers capability at the edge at the cost of 3.76 mW during inference.

Wearable systems for the continuous and real-time monitoring of cardiovascular diseases are becoming widespread and valuable assets in diagnosis and therapy. A promising approach for real-time analysis of the electrocardiographic (ECG) signal and the detection of heart conditions, such as arrhythmia, is represented by the transformer machine learning model. Transformers are powerful models for the classification of time series, although efficient implementation in the wearable domain raises significant design challenges, to combine adequate accuracy and a suitable complexity.In this work, we present a tiny transformer model for the analysis of the ECG signal, requiring only 6k parameters and reaching 98.97% accuracy in the recognition of the 5 most common arrhythmia classes from the MIT-BIH Arrhythmia database, assessed considering 8-bit integer inference as required for efficient execution on low-power microcontroller-based devices. We explored an augmentation-based training approach for improving the robustness against electrode motion artifacts noise, resulting in a worst-case post-deployment performance assessment of 98.36% accuracy. Suitability for wearable monitoring solutions is finally demonstrated through efficient deployment on the parallel ultra-low-power GAP9 processor, where inference execution requires 4.28ms and 0.09mJ.

Spiking Neural Networks (SNNs) are energy-and performance-efficient tools that have been found to be very useful in AI applications at the edge. This paper introduces SYNtzulu, an SNN processing element designed to be used in low-cost and low-power FPGA devices for near-sensor data analysis. The system is equipped with a RISC-V subsystem responsible for controlling the input/output and setting runtime parameters, thus increasing its flexibility. We evaluated the system, which was implemented on a Lattice iCE40UP5K FPGA, in various use cases employing SNNs with accuracy comparable to the state-of-the-art. SYNtzulu dissipates a maximum power of 12.05 mW when performing SNN inference, which can be reduced to an average of just 1.45 mW through the use of dynamic power management.

Classification of surface electromyographic (sEMG) signals for the precise identification of hand gestures is a crucial area in the advancement of complex prosthetic devices and human-machine interfaces. This study presents a real-time sEMG classification system, exploiting a Spiking Neural Network (SNN) to distinguish among twelve distinct hand gestures. The system is implemented on a Lattice iCE40-UltraPlus FPGA, explicitly designed for low-power applications. Evaluation on the NinaPro DB5 dataset confirms an accuracy of 85.6

This study aimed to investigate the presence of structural and functional changes in extraocular muscles (EMs) among patients with inactive Graves’ orbitopathy (GO) classified according to the Clinical Activity Score (CAS). Sixty-seven patients with Graves’ disease (GD) and inactive GO were included. The data collected included clinical parameters, thyroid function, autoantibody levels, EOM morphology via orbital ultrasound (US), and ocular motility. Patients were stratified into Red Filter Test (RFT)-positive or RFT-negative groups based on the presence or absence of latent diplopia during the RFT examination. Thirty-three patients (49.25%) exhibited latent diplopia on the RFT, despite not reporting double vision during standard ocular motility tests. Significant differences were observed between the two groups in terms of age, disease duration, intraocular pressure (IOP) elevation in up-gaze, and medial rectus muscle thickness (p < 0.05). No significant differences were found in thyroid status, TRAb and ATA levels, CASs, exophthalmos, or lateral rectus thickness between the two groups. This study revealed that in inactive GO, subclinical EM dysfunction and morphological changes may be present, which might not be apparent through routine ocular examinations. The RFT is effective in detecting latent diplopia, highlighting its utility in identifying subtle ocular motility issues and subclinical muscle involvement. Comprehensive evaluations combining functional tests like the RFT and imaging are essential for early detection of GO-related abnormalities, enabling tailored and prompt management and improving patient outcomes.

Brain-machine interfaces (BMIs) are systems designed to decode neural signals and translate them into commands for external devices. Intracortical microelectrode arrays (MEAs) represent a significant advancement in this field, offering unprecedented spatial and temporal resolutions for monitoring brain activity. However, processing data from MEAs presents challenges due to high data rates and computing power requirements. To address these challenges, we propose a novel solution leveraging spiking neural networks (SNNs) that, due to their similarity to biological neural networks and their event-based nature, promise high compatibility with neural signals and low energy consumption. In this study, we introduce a real-time neural decoding system based on an SNN, deployed on a Lattice iCE40UP5k FPGA. This system is capable of reconstructing multiple target variables, related to the kinematics and kinetics of hand motion, from iEEG signals recorded by a 96-channel MEA. We evaluated the system using two different public datasets, achieving results similar to state-of-the-art neural decoders that use more complex deep learning models. This was obtained while maintaining an average power consumption of 13.9 mW and an average energy consumption per inference of 13.9 uJ.

In recent years, the rapid evolution of smart technologies has spurred enterprises to undergo digital transformations, revolutionizing their business processes and operations. This shift, known as Digital Transformation, has permeated diverse sectors, particularly impacting production systems. Notably, Artificial Intelligence (AI) and robotic automation have emerged as pivotal drivers in this transformation, promising enhanced efficiency and innovation in industrial digitization. This paper presents a novel architecture designed to facilitate digital transformation within enterprises, harnessing the capabilities of advanced collaborative robots (cobots) and cutting-edge image segmentation techniques. Focused on a practical scenario within a food production environment, our proposed architecture aims to seamlessly integrate a cobot and a camera in an automatic system for efficient cardboard disposal. Specifically, our attention is drawn to the challenge of differentiating sections of food packaging suitable for disposal from those contaminated with stains or organic residues, a task with significant implications for waste management efficiency. By leveraging a cloud-based architecture and deploying AI algorithms for image segmentation, localization, and robot guidance, our study showcases the tangible benefits and practical applicability of these methodologies in real-world settings. This research not only highlights the potential of AI-driven solutions in addressing specific industrial challenges but also underscores the broader impact of digital transformation on optimizing operational processes and driving innovation across sectors.

The identification of cardiac arrhythmias is a significant issue in modern healthcare and a major application for Artificial Intelligence (AI) systems based on artificial neural networks. This research introduces a real-time arrhythmia diagnosis system that uses a Spiking Neural Network (SNN) to classify heartbeats into five types of arrhythmias from a single-lead electrocardiogram (ECG) signal. The system is implemented on a custom SNN processor running on a low-power Lattice iCE40-UltraPlus FPGA. It was tested using the MIT-BIH dataset, and achieved accuracy results that are comparable to the most advanced SNN models, reaching 98.4% accuracy. The proposed modules take advantage of the energy efficiency of SNNs to reduce the average execution time to 4.32 ms and energy consumption to 50.98 uJ per classification.

This paper outlines the initial FPGA-centric endeavors within the EdgeAI project, targeting scenarios where extremely constrained power-energy parameters intersect with the demand for high performance and accuracy in executing Artificial Intelligence (AI) algorithms. Our discussion, after presenting the generalities of the EdgeAI project, revolves around the project objective of leveraging simultaneously event-based spiking neural networks and low-end FPGA chips for very-low-power near-sensor AI inference. We present the hardware/software implementation of this approach and the early results on the project use cases.