The deployment of Deep Neural Networks (DNNs) on end-nodes at the extreme edge of the Internet-of-Things is a critical enabler to support pervasive Deep Learning-enhanced applications. Low-Cost MCU-based end-nodes have limited on-chip memory and often replace caches with scratchpads, to reduce area overheads and increase energy efficiency - requiring explicit DMA-based memory transfers between different levels of the memory hierarchy. Mapping modern DNNs on these systems requires aggressive topology-dependent tiling and double-buffering. In this work, we propose DORY (Deployment Oriented to memoRY) - an automatic tool to deploy DNNs on low cost MCUs with typically less than 1MB of on-chip SRAM memory. DORY abstracts tiling as a Constraint Programming (CP) problem: it maximizes L1 memory utilization under the topological constraints imposed by each DNN layer. Then, it generates ANSI C code to orchestrate off- and on-chip transfers and computation phases. Furthermore, to maximize speed, DORY augments the CP formulation with heuristics promoting performance-effective tile sizes. As a case study for DORY, we target GreenWaves Technologies GAP8, one of the most advanced parallel ultra-low power MCU-class devices on the market. On this device, DORY achieves up to 2.5× better MAC/cycle than the GreenWaves proprietary software solution and 18.1× better than the state-of-the-art result on an STM32-H743 MCU on single layers. Using our tool, GAP-8 can perform end-to-end inference of a 1.0-MobileNet-128 network consuming just 63 pJ/MAC on average @ 4.3 fps - 15.4× better than an STM32-H743. We release all our developments - the DORY framework, the optimized backend kernels, and the related heuristics - as open-source software.

Hand movement classification via surface electromyographic (sEMG) signal is a well-established approach for advanced Human-Computer Interaction. However, sEMG movement recognition has to deal with the long-term reliability of sEMG-based control, limited by the variability affecting the sEMG signal. Embedded solutions are affected by a recognition accuracy drop over time that makes them unsuitable for reliable gesture controller design. In this paper, we present a complete wearable-class embedded system for robust sEMG-based gesture recognition, based on Temporal Convolutional Networks (TCNs). Firstly, we developed a novel TCN topology (TEMPONet), and we tested our solution on a benchmark dataset (Ninapro), achieving 49.6% average accuracy, 7.8%, better than current State-Of-the-Art (SoA). Moreover, we designed an energy-efficient embedded platform based on GAP8, a novel 8-core IoT processor. Using our embedded platform, we collected a second 20-sessions dataset to validate the system on a setup which is representative of the final deployment. We obtain 93.7% average accuracy with the TCN, comparable with a SoA SVM approach (91.1%). Finally, we profiled the performance of the network implemented on GAP8 by using an 8-bit quantization strategy to fit the memory constraint of the processor. We reach a 4× lower memory footprint (460 kB) with a performance degradation of only 3% accuracy. We detailed the execution on the GAP8 platform, showing that the quantized network executes a single classification in 12.84 ms with a power envelope of 0.9 mJ, making it suitable for a long-lifetime wearable deployment.

Objective: We develop a fast learning algorithm combining symbolic dynamics and brain-inspired hyperdimensional computing for both seizure onset detection and identification of ictogenic (seizure generating) brain regions from intracranial electroencephalography (iEEG). Methods: Our algorithm first transforms iEEG time series from each electrode into symbolic local binary pattern codes from which a holographic distributed representation of the brain state of interest is constructed across all the electrodes and over time in a hyperdimensional space. The representation is used to quickly learn from few seizures, detect their onset, and identify the spatial brain regions that generated them. Results: We assess our algorithm on our dataset that contains 99 short-time iEEG recordings from 16 drug-resistant epilepsy patients being implanted with 36 to 100 electrodes. For the majority of the patients (10 out of 16), our algorithm quickly learns from one or two seizures and perfectly (100%) generalizes on novel seizures using k-fold cross-validation. For the remaining six patients, the algorithm requires three to six seizures for learning. Our algorithm surpasses the state-of-the-art including deep learning algorithms by achieving higher specificity (94.84% vs. 94.77%) and macroaveraging accuracy (95.42% vs. 94.96%), and 74x lower memory footprint, but slightly higher average latency in detection (15.9 s vs. 14.7 s). Moreover, the algorithm can reliably identify (with a p-value < 0.01) the relevant electrodes covering an ictogenic brain region at two levels of granularity: cerebral hemispheres and lobes. Conclusion and significance: Our algorithm provides: (1) a unified method for both learning and classification tasks with end-to-end binary operations; (2) one-shot learning from seizure examples; (3) linear computational scalability for increasing number of electrodes; (4) generation of transparent codes that enables post-translational supports for clinical decision making.

Many emerging applications of nano-sized unmanned aerial vehicles (UAVs), with a few cm² form-factor, revolve around safely interacting with humans in complex scenarios, for example, monitoring their activities or looking after people needing care. Such sophisticated autonomous functionality must be achieved while dealing with severe constraints in payload, battery, and power budget (~100mW). In this work, we attack a complex task going from perception to control: to estimate and maintain the nano-UAV’s relative 3-D pose with respect to a person while they freely move in the environment—a task that, to the best of our knowledge, has never previously been targeted with fully onboard computation on a nano-sized UAV. Our approach is centered around a novel vision-based deep neural network (DNN), called Frontnet, designed for deployment on top of a parallel ultra-low power (PULP) processor aboard a nano-UAV. We present a vertically integrated approach starting from the DNN model design, training, and dataset augmentation down to 8-bit quantization and deployment in-field. PULP-Frontnet can operate in real-time (up to 135 frame/s), consuming less than 87 mW for processing at peak throughput and down to 0.43 mJ/frame in the most energy-efficient operating point. Field experiments demonstrate a closed-loop top-notch autonomous navigation capability, with a tiny 27-g Crazyflie 2.1 nano-UAV. Compared against an ideal sensing setup, onboard pose inference yields excellent drone behavior in terms of median absolute errors, such as positional (onboard: 41cm, ideal: 26 cm) and angular (onboard: 3.7°, ideal: 4.1°). We publicly release videos and the source code of our work.

Context: Primary aldosteronism (PA) comprises unilateral (lateralized [LPA]) and bilateral disease (BPA). The identification of LPA is important to recommend potentially curative adrenalectomy. Adrenal venous sampling (AVS) is considered the gold standard for PA subtyping, but the procedure is available in few referral centers. Objective: To develop prediction models for subtype diagnosis of PA using patient clinical and biochemical characteristics. Design, Patients and Setting: Patients referred to a tertiary hypertension unit. Diagnostic algorithms were built and tested in a training (N = 150) and in an internal validation cohort (N = 65), respectively. The models were validated in an external independent cohort (N = 118). Main outcome measure: Regression analyses and supervised machine learning algorithms were used to develop and validate 2 diagnostic models and a 20-point score to classify patients with PA according to subtype diagnosis. Results: Six parameters were associated with a diagnosis of LPA (aldosterone at screening and after confirmatory testing, lowest potassium value, presence/absence of nodules, nodule diameter, and computed tomography results) and were included in the diagnostic models. Machine learning algorithms displayed high accuracy at training and internal validation (79.1%-93%), whereas a 20-point score reached an area under the curve of 0.896, and a sensitivity/specificity of 91.7/79.3%. An integrated flowchart correctly addressed 96.3% of patients to surgery and would have avoided AVS in 43.7% of patients. The external validation on an independent cohort confirmed a similar diagnostic performance. Conclusions: Diagnostic modelling techniques can be used for subtype diagnosis and guide surgical decision in patients with PA in centers where AVS is unavailable.

Hearth Rate (HR) monitoring is increasingly performed in wrist-worn devices using low-cost photoplethysmography (PPG) sensors. However, Motion Artifacts (MAs) caused by movements of the subject’s arm affect the performance of PPG-based HR tracking. This is typically addressed coupling the PPG signal with acceleration measurements from an inertial sensor. Unfortunately, most standard approaches of this kind rely on hand-tuned parameters, which impair their generalization capabilities and their applicability to real data in the field. In contrast, methods based on deep learning, despite their better generalization, are considered to be too complex to deploy on wearable devices. In this work, we tackle these limitations, proposing a design space exploration methodology to automatically generate a rich family of deep Temporal Convolutional Networks (TCNs) for HR monitoring, all derived from a single “seed” model. Our flow involves a cascade of two Neural Architecture Search (NAS) tools and a hardware-friendly quantizer, whose combination yields both highly accurate and extremely lightweight models. When tested on the PPG-Dalia dataset, our most accurate model sets a new state-of-the-art in Mean Absolute Error. Furthermore, we deploy our TCNs on an embedded platform featuring a STM32WB55 microcontroller, demonstrating their suitability for real-time execution. Our most accurate quantized network achieves 4.41 Beats Per Minute (BPM) of Mean Absolute Error (MAE), with an energy consumption of 47.65 mJ and a memory footprint of 412 kB. At the same time, the smallest network that obtains a MAE $< $ 8 BPM, among those generated by our flow, has a memory footprint of 1.9 kB and consumes just 1.79 mJ per inference.