Publications related to Laboratoire de systemes microelectroniques

Challenges and Opportunities of Edge AI for Next-Generation Implantable BMIs

Neuroscience and neurotechnology are currently being revolutionized by artificial intelligence (AI) and machine learning. AI is widely used to study and interpret neural signals (analytical applications), assist people with disabilities (prosthetic applications), and treat underlying neurological symptoms (therapeutic applications). In this brief, we will review the emerging opportunities of on-chip AI for the next-generation implantable brain machine interfaces (BMIs), with a focus on state-of-the-art prosthetic BMIs. Major technological challenges for the effectiveness of AI models will be discussed. Finally, we will present algorithmic and IC design solutions to enable a new generation of AI-enhanced and high-channel-count BMIs.

IEEE2022

,

The researchers used a machine-learning classification approach to better understand neurological features associated with periods of wayfinding uncertainty. The participants (n = 30) were asked to complete wayfinding tasks of varying difficulty in a virtual reality (VR) hospital environment. Time segments when participants experienced navigational uncertainty were first identified using a combination of objective measurements (frequency of inputs into the VR controller) and behavioral annotations from two independent observers. Uncertainty time-segments during navigation were ranked on a scale from 1 (low) to 5 (high). The machine-learning model, a Random Forest classifier implemented using scikit-learn in Python, was used to evaluate common spatial patterns of EEG spectral power across the theta, alpha, and beta bands associated with the researcher-identified uncertainty states. The overall predictive power of the resulting model was 0.70 in terms of the area under the Receiver Operating Characteristics curve (ROC-AUC). These findings indicate that EEG data can potentially be used as a metric for identifying navigational uncertainty states, which may provide greater rigor and efficiency in studies of human responses to architectural design variables and wayfinding cues.

ELSEVIER SCI LTD2022

Epileptic Seizure Detection With Patient-Specific Feature and Channel Selection for Low-power Applications

Keyvan Farhang Razi, Alexandre Schmid

An accurate epileptic seizure detector using intracranial electroencephalography (iEEG) recordings, suitable for low-power wearable/implantable applications, is presented. Eleven time-domain features with low hardware complexity are employed in the feature pool of the seizure detector. A novel two-step feature ranking algorithm based on maximum discrimination minimum redundancy (MDMR) is proposed to identify the most discriminating features in a patient-specific manner during a training phase. Subsequently, the top-ranked features are extracted in a two-stage energy-efficient architecture in which the second stage is activated by a controller to avoid unnecessary energy consumption imposed by multiple feature extraction during the long period of non-seizure states. The effective number of features which are continuously extracted is reduced in this work. Moreover, a patient-specific data reduction technique which selects the most informative and discriminating iEEG channels is also proposed. The presented channel selection technique reaches 68% data reduction on average for the tested patients and reduces the computational complexity. The proposed algorithm is implemented on Cyclone V FPGA of Terasic DE10-standard board. It is tested on twelve patients with short-term and long-term seizures from the Bern Inselspital dataset. The FPGA implementation results reveal an excellent sensitivity of 100% for all patients and remarkable specificity and detection delay improvement compared to the state-of-the-art. The average dynamic power consumption is 0.49 mW which is in the acceptable range for low-power wearable/implantable applications. In addition, a new figure-of-merit (FoM_SD) is defined to consider the major parameters of the seizure detectors. The outstanding FoM_SD of this work is 0.464 which outperforms the state-of-the-art.

IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC2022

Effect of Native Oxide on Stress in Silicon Nanowires: Implications for Nanoelectromechanical Systems

Yusuf Leblebici, Mohammad Nasr Esfahani, Zuhal Tasdemir

Understanding the origins of intrinsic stress in Si nanowires (NWs) is crucial for their successful utilization as transducer building blocks in next-generation, miniaturized sensors based on nanoelectromechanical systems (NEMS). With their small size leading to ultrahigh-resonance frequencies and extreme surface-to-volume ratios, silicon NWs raise new opportunities regarding sensitivity, precision, and speed in both physical and biochemical sensing. With silicon optoelectromechanical properties strongly dependent on the level of NW intrinsic stress, various studies have been devoted to the measurement of such stresses generated, for example, as a result of harsh fabrication processes. However, due to enormous NW surface area, even the native oxide that is conventionally considered as a benign surface condition can cause significant stresses. To address this issue, a combination of nanomechanical characterization and atomistic simulation approaches is developed. Relying only on low-temperature processes, the fabrication approach yields monolithic NWs with optimum boundary conditions, where NWs and support architecture are etched within the same silicon crystal. Resulting NWs are characterized by transmission electron microscopy and micro-Raman spectroscopy. The interpretation of results is carried out through molecular dynamics simulations with ReaxFF potential facilitating the incorporation of humidity and temperature, thereby providing a close replica of the actual oxidation environment-in contrast to previous dry oxidation or self-limiting thermal oxidation studies. As a result, consensus on significant intrinsic tensile stresses on the order of 100 MPa to 1 GPa was achieved as a function of NW critical dimension and aspect ratio. The understanding developed herein regarding the role of native oxide played in the generation of NW intrinsic stresses is important for the design and development of silicon-based NEMS.

AMER CHEMICAL SOC2022

Proximal Point Imitation Learning

Volkan Cevher, Angeliki Kamoutsi, Igor Krawczuk, Luca Viano

This work develops new algorithms with rigorous efficiency guarantees for infinite horizon imitation learning (IL) with linear function approximation without restrictive coherence assumptions. We begin with the minimax formulation of the problem and then outline how to leverage classical tools from optimization, in particular, the proximal-point method (PPM) and dual smoothing, for online and offline IL, respectively. Thanks to PPM, we avoid nested policy evaluation and cost updates for online IL appearing in the prior literature. In particular, we do away with the conventional alternating updates by the optimization of a single convex and smooth objective over both cost and Q-functions. When solved inexactly, we relate the optimization errors to the suboptimality of the recovered policy. As an added bonus, by re-interpreting PPM as dual smoothing with the expert policy as a center point, we also obtain an offline IL algorithm enjoying theoretical guarantees in terms of required expert trajectories. Finally, we achieve convincing empirical performance for both linear and neural network function approximation.

2022

Systems and methods for energy-efficient measurement of neurophysiological oscillations

Uisub Shin, Mahsa Shoaran

A system for measuring synchrony between two or more regions, or within a single region, of a subject's brain. The system can include a signal conditioning module in communication with a phase extraction module. The signal conditioning module can receive an oscillating signal, bandpass filter the signal, and convert the signal to its real and imaginary components. The phase extraction module can receive the real and imaginary components, determine the quadrant of the oscillating signal on the complex plane based on the real and imaginary components, and determine the phase of the oscillating signal using a linear arctangent approximation based on the determined quadrant. The system can calculate synchrony metrics based on the determined phase.

2022

Hardware-oriented pruning and quantization of Deep Learning models to detect life-threatening arrhythmias

Lizeth Gonzalez Carabarin, Alexandre Schmid

Wearable solutions based on Deep Learning (DL) for real-time ECG monitoring are a promising alternative to detect life-threatening arrhythmias. However, DL models suffer of a large memory footprint, which hampers their adoption in portable technologies. Therefore, we leverage a hardware-oriented pruning approach to effectively shrink DL models. We demonstrate that tiny DL models can be reduced to 5.55x (pruning), and 26.6x (pruning+quantization) compression rate, with 82.9% FLOP's reduction. These ultra-compressed models are able to effectively classify life-threatening arrhythmias with minimal or no loss of performance compared with their non=pruned counterparts, which can pave the path towards DL-based biomedical portable solutions.

IEEE2021

Dynamic Probabilistic Pruning: A General Framework for Hardware-Constrained Pruning at Different Granularities

Lizeth Gonzalez Carabarin, Alexandre Schmid

Unstructured neural network pruning algorithms have achieved impressive compression ratios. However, the resulting-typically irregular-sparse matrices hamper efficient hardware implementations, leading to additional memory usage and complex control logic that diminishes the benefits of unstructured pruning. This has spurred structured coarse-grained pruning solutions that prune entire feature maps or even layers, enabling efficient implementation at the expense of reduced flexibility. Here, we propose a flexible new pruning mechanism that facilitates pruning at different granularities (weights, kernels, and feature maps) while retaining efficient memory organization (e.g., pruning exactly k-out-of- n weights for every output neuron or pruning exactly k-out-of-n kernels for every feature map). We refer to this algorithm as dynamic probabilistic pruning (DPP). DPP leverages the Gumbel-softmax relaxation for differentiable k-out-of-n sampling, facilitating end-to-end optimization. We show that DPP achieves competitive compression ratios and classification accuracy when pruning common deep learning models trained on different benchmark datasets for image classification. Relevantly, the dynamic masking of DPP facilitates for joint optimization of pruning and weight quantization in order to even further compress the network, which we show as well. Finally, we propose novel information-theoretic metrics that show the confidence and pruning diversity of pruning masks within a layer.

IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC2022

Resonance Frequency Dependence Of A1 Lamb Mode On The Pitch Of The Electrode Structure

Seniz Esra Küçük, Luis Guillermo Villanueva Torrijo, Soumya Yandrapalli

Dispersion of A1 Lamb mode near onset frequency is discussed. This mode is close to that used in XBAR devices. The typically used approximation is based on elliptic form of slowness curve and ignoring mode interaction is not valid for widely used materials, such as ZY-LiNbO3. Alternative phenomenological dispersion formula is proposed.

IEEE2021

,

A novel low-complexity method of detecting epileptic seizures from intracranial encephalography (iEEG) signals is presented. In the proposed algorithm, coastline, energy and nonlinear energy features of iEEG signals are extracted in a patient-specific two-stage seizure detection system. The detection stage of the proposed system, which extracts two times more features than the monitoring stage, is only powered on when the monitoring stage detects a seizure occurrence. A new metric is defined to demonstrate the significance of the two-stage architecture and show the time duration over which the detection stage is activated. The new parameter is called detection stage activation ratio (DAR) and it is equal to 0.272 in this work. In addition, the proposed seizure detector outperforms other algorithms which utilize a single feature or multiple features continuously in terms of sensitivity, specificity and DAR. Therefore, it is highly suitable for seizure detector implants in which reducing the power consumption is a critical factor to increase the lifetime of the implanted battery. The algorithm is implemented on a Cyclone V FPGA and has a low dynamic power of 1 mu W when tested on human iEEG signals of six patients from the Bern Inselspital dataset. It reaches a perfect sensitivity of 100% tested on 120 hours of iEEG data containing 24 seizure periods of six patients.

IEEE2021