A compute-in-memory chip based on resistive random-access memory.

Realizing increasingly complex artificial intelligence (AI) functionalities directly on edge devices calls for unprecedented energy efficiency of edge hardware. Compute-in-memory (CIM) based on resistive random-access memory (RRAM)1 promises to meet such demand by storing AI model weights in dense, analogue and non-volatile RRAM devices, and by performing AI computation directly within RRAM, thus eliminating power-hungry data movement between separate compute and memory2-5. Although recent studies have demonstrated in-memory matrix-vector multiplication on fully integrated RRAM-CIM hardware6-17, it remains a goal for a RRAM-CIM chip to simultaneously deliver high energy efficiency, versatility to support diverse models and software-comparable accuracy. Although efficiency, versatility and accuracy are all indispensable for broad adoption of the technology, the inter-related trade-offs among them cannot be addressed by isolated improvements on any single abstraction level of the design. Here, by co-optimizing across all hierarchies of the design from algorithms and architecture to circuits and devices, we present NeuRRAM-a RRAM-based CIM chip that simultaneously delivers versatility in reconfiguring CIM cores for diverse model architectures, energy efficiency that is two-times better than previous state-of-the-art RRAM-CIM chips across various computational bit-precisions, and inference accuracy comparable to software models quantized to four-bit weights across various AI tasks, including accuracy of 99.0 percent on MNIST18 and 85.7 percent on CIFAR-1019 image classification, 84.7-percent accuracy on Google speech command recognition20, and a 70-percent reduction in image-reconstruction error on a Bayesian image-recovery task.

Spatial encoding in primate hippocampus during free navigation.

The hippocampus comprises two neural signals-place cells and θ oscillations-that contribute to facets of spatial navigation. Although their complementary relationship has been well established in rodents, their respective contributions in the primate brain during free navigation remains unclear. Here, we recorded neural activity in the hippocampus of freely moving marmosets as they naturally explored a spatial environment to more explicitly investigate this issue. We report place cells in marmoset hippocampus during free navigation that exhibit remarkable parallels to analogous neurons in other mammalian species. Although θ oscillations were prevalent in the marmoset hippocampus, the patterns of activity were notably different than in other taxa. This local field potential oscillation occurred in short bouts (approximately .4 s)-rather than continuously-and was neither significantly modulated by locomotion nor consistently coupled to place-cell activity. These findings suggest that the relationship between place-cell activity and θ oscillations in primate hippocampus during free navigation differs substantially from rodents and paint an intriguing comparative picture regarding the neural basis of spatial navigation across mammals.

A Neuromorphic Brain Interface based on RRAM Crossbar Arrays for High Throughput Real-time Spike Sorting.

Real-time spike sorting and processing are crucial for closed-loop brain-machine interfaces and neural prosthetics. Recent developments in high-density multi-electrode arrays with hundreds of electrodes have enabled simultaneous recordings of spikes from a large number of neurons. However, the high channel count imposes stringent demands on real-time spike sorting hardware regarding data transmission bandwidth and computation complexity. Thus, it is necessary to develop a specialized real-time hardware that can sort neural spikes on the fly with high throughputs while consuming minimal power. Here, we present a real-time, low latency spike sorting processor that utilizes high-density CuOx resistive crossbars to implement in-memory spike sorting in a massively parallel manner. We developed a fabrication process which is compatible with CMOS BEOL integration. We extensively characterized switching characteristics and statistical variations of the CuOx memory devices. In order to implement spike sorting with crossbar arrays, we developed a template matching-based spike sorting algorithm that can be directly mapped onto RRAM crossbars. By using synthetic and in vivo recordings of extracellular spikes, we experimentally demonstrated energy efficient spike sorting with high accuracy. Our neuromorphic interface offers substantial improvements in area (~1000× less area), power (~200× less power), and latency (4.8µs latency for sorting 100 channels) for real-time spike sorting compared to other hardware implementations based on FPGAs and microcontrollers.

Array atomic force microscopy for real-time multiparametric analysis.

Nanoscale multipoint structure-function analysis is essential for deciphering the complexity of multiscale biological and physical systems. Atomic force microscopy (AFM) allows nanoscale structure-function imaging in various operating environments and can be integrated seamlessly with disparate probe-based sensing and manipulation technologies. Conventional AFMs only permit sequential single-point analysis; widespread adoption of array AFMs for simultaneous multipoint study is challenging owing to the intrinsic limitations of existing technological approaches. Here, we describe a prototype dispersive optics-based array AFM capable of simultaneously monitoring multiple probe-sample interactions. A single supercontinuum laser beam is utilized to spatially and spectrally map multiple cantilevers, to isolate and record beam deflection from individual cantilevers using distinct wavelength selection. This design provides a remarkably simplified yet effective solution to overcome the optical cross-talk while maintaining subnanometer sensitivity and compatibility with probe-based sensors. We demonstrate the versatility and robustness of our system on parallel multiparametric imaging at multiscale levels ranging from surface morphology to hydrophobicity and electric potential mapping in both air and liquid, mechanical wave propagation in polymeric films, and the dynamics of living cells. This multiparametric, multiscale approach provides opportunities for studying the emergent properties of atomic-scale mechanical and physicochemical interactions in a wide range of physical and biological networks.

Combining biophysical modeling and deep learning for multielectrode array neuron localization and classification.

Neural circuits typically consist of many different types of neurons, and one faces a challenge in disentangling their individual contributions in measured neural activity. Classification of cells into inhibitory and excitatory neurons and localization of neurons on the basis of extracellular recordings are frequently employed procedures. Current approaches, however, need a lot of human intervention, which makes them slow, biased, and unreliable. In light of recent advances in deep learning techniques and exploiting the availability of neuron models with quasi-realistic three-dimensional morphology and physiological properties, we present a framework for automatized and objective classification and localization of cells based on the spatiotemporal profiles of the extracellular action potentials recorded by multielectrode arrays. We train convolutional neural networks on simulated signals from a large set of cell models and show that our framework can predict the position of neurons with high accuracy, more precisely than current state-of-the-art methods. Our method is also able to classify whether a neuron is excitatory or inhibitory with very high accuracy, substantially improving on commonly used clustering techniques. Furthermore, our new method seems to have the potential to separate certain subtypes of excitatory and inhibitory neurons. The possibility of automatically localizing and classifying all neurons recorded with large high-density extracellular electrodes contributes to a more accurate and more reliable mapping of neural circuits. NEW & NOTEWORTHY We propose a novel approach to localize and classify neurons from their extracellularly recorded action potentials with a combination of biophysically detailed neuron models and deep learning techniques. Applied to simulated data, this new combination of forward modeling and machine learning yields higher performance compared with state-of-the-art localization and classification methods.

A Learning Framework for Winner-Take-All Networks with Stochastic Synapses.

Many recent generative models make use of neural networks to transform the probability distribution of a simple low-dimensional noise process into the complex distribution of the data. This raises the question of whether biological networks operate along similar principles to implement a probabilistic model of the environment through transformations of intrinsic noise processes. The intrinsic neural and synaptic noise processes in biological networks, however, are quite different from the noise processes used in current abstract generative networks. This, together with the discrete nature of spikes and local circuit interactions among the neurons, raises several difficulties when using recent generative modeling frameworks to train biologically motivated models. In this letter, we show that a biologically motivated model based on multilayer winner-take-all circuits and stochastic synapses admits an approximate analytical description. This allows us to use the proposed networks in a variational learning setting where stochastic backpropagation is used to optimize a lower bound on the data log likelihood, thereby learning a generative model of the data. We illustrate the generality of the proposed networks and learning technique by using them in a structured output prediction task and a semisupervised learning task. Our results extend the domain of application of modern stochastic network architectures to networks where synaptic transmission failure is the principal noise mechanism.

A Low-Noise Low-Power 0.001Hz-1kHz Neural Recording System-on-Chip With Sample-Level Duty-Cycling.

Advances in brain-machine interfaces and wearable biomedical sensors for healthcare and human-computer interactions call for precision electrophysiology to resolve a variety of biopotential signals across the body that cover a wide range of frequencies, from the mHz-range electrogastrogram (EGG) to the kHz-range electroneurogram (ENG). Existing integrated wearable solutions for minimally invasive biopotential recordings are limited in detection range and accuracy due to trade-offs in bandwidth, noise, input impedance, and power consumption. This article presents a 16-channel wide-band ultra-low-noise neural recording system-on-chip (SoC) fabricated in 65nm CMOS for chronic use in mobile healthcare settings that spans a bandwidth of 0.001 Hz to 1 kHz through a featured sample-level duty-cycling (SLDC) mode. Each recording channel is implemented by a delta-sigma analog-to-digital converter (ADC) achieving 1.0 µ V rms input-referred noise over 1Hz-1kHz bandwidth with a Noise Efficiency Factor (NEF) of 2.93 in continuous operation mode. In SLDC mode, the power supply is duty-cycled while maintaining consistently low input-referred noise levels at ultra-low frequencies (1.1 µV rms over 0.001Hz-1Hz) and 435 M Ω input impedance. The functionalities of the proposed SoC are validated with two human electrophysiology applications: recording low-amplitude electroencephalogram (EEG) through electrodes fixated on the forehead to monitor brain waves, and ultra-slow-wave electrogastrogram (EGG) through electrodes fixated on the abdomen to monitor digestion.

Multi-level, forming and filament free, bulk switching trilayer RRAM for neuromorphic computing at the edge.

CMOS-RRAM integration holds great promise for low energy and high throughput neuromorphic computing. However, most RRAM technologies relying on filamentary switching suffer from variations and noise, leading to computational accuracy loss, increased energy consumption, and overhead by expensive program and verify schemes. We developed a filament-free, bulk switching RRAM technology to address these challenges. We systematically engineered a trilayer metal-oxide stack and investigated the switching characteristics of RRAM with varying thicknesses and oxygen vacancy distributions to achieve reliable bulk switching without any filament formation. We demonstrated bulk switching at megaohm regime with high current nonlinearity, up to 100 levels without compliance current. We developed a neuromorphic compute-in-memory platform and showcased edge computing by implementing a spiking neural network for an autonomous navigation/racing task. Our work addresses challenges posed by existing RRAM technologies and paves the way for neuromorphic computing at the edge under strict size, weight, and power constraints.

A Versatile In-Ear Biosensing System and Body-Area Network for Unobtrusive Continuous Health Monitoring.

To enable continuous, mobile health monitoring, body-worn sensors need to offer comparable performance to clinical devices in a lightweight, unobtrusive package. This work presents a complete versatile wireless electrophysiology data acquisition system (weDAQ) that is demonstrated for in-ear electroencephalography (EEG) and other on-body electrophysiology with user-generic dry-contact electrodes made from standard printed circuit boards (PCBs). Each weDAQ device provides 16 recording channels, driven right leg (DRL), a 3-axis accelerometer, local data storage, and adaptable data transmission modes. The weDAQ wireless interface supports deployment of a body area network (BAN) capable of aggregating various biosignal streams over multiple worn devices simultaneously, on the 802.11n WiFi protocol. Each channel resolves biopotentials ranging over 5 orders of magnitude with a noise level of 0.52 µVrms over a 1000-Hz bandwidth, and a peak SNDR of 119 dB and CMRR of 111 dB at 2 ksps. The device leverages in-band impedance scanning and an input multiplexer to dynamically select good skin contacting electrodes for reference and sensing channels. In-ear and forehead EEG measurements taken from subjects captured modulation of alpha brain activity, electrooculogram (EOG) characteristic eye movements, and electromyogram (EMG) from jaw muscles. Simultaneous ECG and EMG measurements were demonstrated on multiple, freely-moving subjects in their natural office environment during periods of rest and exercise. The small footprint, performance, and configurability of the open-source weDAQ platform and scalable PCB electrodes presented, aim to provide the biosensing community greater experimental flexibility and lower the barrier to entry for new health monitoring research.

A 22-pJ/spike 73-Mspikes/s 130k-compartment neural array transceiver with conductance-based synaptic and membrane dynamics.

Neuromorphic cognitive computing offers a bio-inspired means to approach the natural intelligence of biological neural systems in silicon integrated circuits. Typically, such circuits either reproduce biophysical neuronal dynamics in great detail as tools for computational neuroscience, or abstract away the biology by simplifying the functional forms of neural computation in large-scale systems for machine intelligence with high integration density and energy efficiency. Here we report a hybrid which offers biophysical realism in the emulation of multi-compartmental neuronal network dynamics at very large scale with high implementation efficiency, and yet with high flexibility in configuring the functional form and the network topology. The integrate-and-fire array transceiver (IFAT) chip emulates the continuous-time analog membrane dynamics of 65 k two-compartment neurons with conductance-based synapses. Fired action potentials are registered as address-event encoded output spikes, while the four types of synapses coupling to each neuron are activated by address-event decoded input spikes for fully reconfigurable synaptic connectivity, facilitating virtual wiring as implemented by routing address-event spikes externally through synaptic routing table. Peak conductance strength of synapse activation specified by the address-event input spans three decades of dynamic range, digitally controlled by pulse width and amplitude modulation (PWAM) of the drive voltage activating the log-domain linear synapse circuit. Two nested levels of micro-pipelining in the IFAT architecture improve both throughput and efficiency of synaptic input. This two-tier micro-pipelining results in a measured sustained peak throughput of 73 Mspikes/s and overall chip-level energy efficiency of 22 pJ/spike. Non-uniformity in digitally encoded synapse strength due to analog mismatch is mitigated through single-point digital offset calibration. Combined with the flexibly layered and recurrent synaptic connectivity provided by hierarchical address-event routing of registered spike events through external memory, the IFAT lends itself to efficient large-scale emulation of general biophysical spiking neural networks, as well as rate-based mapping of rectified linear unit (ReLU) neural activations.

Stimulus Design for Visual Evoked Potential Based Brain-Computer Interfaces.

Visual stimuli design plays an important role in brain-computer interfaces (BCIs) based on visual evoked potentials (VEPs). Variations in stimulus parameters have been shown to affect both decoding accuracy and subjective perception experience, implying the need for a trade-off in design. In this study, we comprehensively and systematically compared various combinations of amplitude contrast and spectral content parameters in the stimulus design to quantify their impact on decoding performance and subject comfort. Specifically, three parameters were investigated: 1) contrast level, 2) temporal pattern (periodic steady-state or pseudo-random code-modulated), and 3) frequency range. We collected electroencephalogram (EEG) data and subjective perception ratings from ten subjects and evaluated the decoding accuracy and subject comfort rating for different combinations of the stimulus parameters. Our results indicate that while high-frequency steady-state VEP (SSVEP) stimuli were rated the most comfortable, they also had the lowest decoding accuracy. Conversely, low-frequency SSVEP stimuli were rated the least comfortable but had the highest decoding accuracy. Standard and high-frequency M-sequence code-modulated VEPs (c-VEPs) produced intermediates between the two. We observed a consistent trade-off relationship between decoding accuracy and subjective comfort level across all parameters. Based on our findings, we offer c-VEP as a preferable stimulus for achieving reliable decoding accuracy while maintaining a reasonable level of comfortability.

Quantitative Simulation of Enzymatic Breakdown of Alcohol in Human Metabolism.

The breakdown of ethanol, the active chemical in alcohol, is tightly regulated by the body, yet alcohol intoxication occurs in thousands of Americans annually. Many factors contribute to the concentration of ethanol in the bloodstream and the tolerance an individual has, including body size, previous drinking experience, and liver functionality. We propose a model that estimates both the blood alcohol concentration and the concentration of acetaldehyde (the toxic intermediate during catabolism) in the liver over time to quantify organ damage for an average person. From the current literature, we derived ordinary differential equations that govern the absorption of ethanol in the body and extended it with the metabolic enzyme mechanisms. We also altered the parameters of our system in order to show the effects of Asian flush, which impairs the body's processing of acetaldehyde. We demonstrated the accumulation of acetaldehyde in Asian flush patients was about 660 times higher compared to those without the disease.Clinical relevance-With further improvements and personalization, our model would be able to quantitatively describe the effects of alcohol consumption without having volunteers go through repetitive trials with extensive exposure to alcohol. Liver damage can also be estimated with the acetaldehyde buildup predicted by the model.

Scalable Anatomically-Tunable Fully In-Ear Dry-Electrode Array for User-Generic Unobtrusive Electrophysiology.

Traditional scalp EEG instrumentation is bulky and arduous to set up, requiring wires that constrain the subject's movement, conductive wet gels that dry over time which limits long-term recording, and/or is socially stigmatized. Therefore, there is growing research in in-ear EEG to increase user wearability, ease of use, and concealability. However, the fabrication of in-ear EEG sensors utilizes complex equipment and materials to capture the intricate geometry of the ear and to fabricate custom earpieces and electrodes. This work aims to lower the barrier of entry by decreasing the fabrication complexity by using PCB components with versatile, user-generic designs. Measured results on the assembled earpiece demonstrate that it viably captures eye blinks, jaw clench, auditory steady-state response (ASSR), and alpha modulation. Additionally, electrochemical impedance spectroscopy (EIS) experiments show reliable electrode-skin contact with impedance comparable to conventional dry-electrode designs at substantially greater channel density.

A Machine Learning Approach to COVID-19 Detection via Graphene Field-Effect-Transistor (GFET).

In the wake of the COVID-19 pandemic, there has been a need for reliable diagnostic testing. However, state-of-the-art detection methods rely on laboratory tests and also vary in accuracy. We evaluate that the usage of a graphene field-effect-transistor (GFET) coupled with machine learning can be a promising alternate diagnostic testing method. We processed the current-voltage data gathered from the GFET sensors to assess information about the presence of COVID-19 in biosamples. We perform binary classification using the following machine learning algorithms: Linear Discriminant Analysis (LDA), Support Vector Machines (SVM) with the Radial Basis Function (RBF) kernel, and K-Nearest Neighbors (KNN) in conjunction with Principal Component Analysis (PCA). We find that LDA and SVM with RBF proved to be the most accurate in identifying positive and negative samples, with accuracies of 99% and 98.5%, respectively. Based on these results, there is promise to develop a bioelectronic diagnostic method for COVID-19 detection by combining GFET technology with machine learning.

In-ear integrated sensor array for the continuous monitoring of brain activity and of lactate in sweat.

Owing to the proximity of the ear canal to the central nervous system, in-ear electrophysiological systems can be used to unobtrusively monitor brain states. Here, by taking advantage of the ear's exocrine sweat glands, we describe an in-ear integrated array of electrochemical and electrophysiological sensors placed on a flexible substrate surrounding a user-generic earphone for the simultaneous monitoring of lactate concentration and brain states via electroencephalography, electrooculography and electrodermal activity. In volunteers performing an acute bout of exercise, the device detected elevated lactate levels in sweat concurrently with the modulation of brain activity across all electroencephalography frequency bands. Simultaneous and continuous unobtrusive in-ear monitoring of metabolic biomarkers and brain electrophysiology may allow for the discovery of dynamic and synergetic interactions between brain and body biomarkers in real-world settings for long-term health monitoring or for the detection or monitoring of neurodegenerative diseases.

Modeling the Viral Kinetics of Influenza A During Infection in Humans.

This study explores the natural control system in the body for responding to exposure to the Influenza A virus. More specifically, it delves into the development of a model to simulate the responses of target uninfected cell counts, infected cell counts, and viral titers. There are two particular models of interest: a delayed model that incorporates the brief inactive period for newly infected cells, and a non-delayed model reflecting only infected cells without delay after initial infection. Both models are commonly used in the literature and the benefits of each model are studied and explained. We generate Simulink models for both the delayed and non-delayed sets of ordinary differential equations (ODEs) to simulate responses to different viral titer impulses. Additionally, this study aims to extrapolate these models to the case for a vaccinated individual. To do this, we modify the viral clearance rate and infected cell death rate of our initial model to account for the improved immune response generated by vaccines.

Luteinizing Hormone Dynamics in Menstruation.

Menstruation is a finely-controlled cycle that responds to the prevailing endocrine and paracrine environment. However, social stigma has led to inadequate menstrual literacy, both among academics and the larger public. The poorly understood mechanisms of menstruation ultimately lead to suboptimal healthcare treatment and services for biological females, culminating in a physical, financial, and emotional burden. Various hormones signal the beginning and end of each stage of menstruation. In particular, luteinizing hormone (LH) is a major player in ovulation, corpus luteum function, and the stimulation of other key hormones. A LH model could be used to understand the larger control system of menstruation if analyzed in conjunction with models for other major hormones (e.g., FSH, progesterone, and GnRH). Thus, exploring a smaller subsection of LH dynamics within the larger control system of menstruation can lead to a greater understanding of menstruation, contributing towards therapeutics and research for women's health. Using parameters and kinetic equations in the existing body of literature, a transfer function was derived to model LH dynamics. Analysis of system stability reveals overdamped dynamics in LH sensitization at baseline, and underdamped mildly resonant dynamics at the peak of the menstrual cycle, the strength of which depends on the values of the rate constants of LH receptor formation, binding, and desensitization.

Unity Human Eye Model for Gaze Tracking with a Query-Driven Dynamic Vision Sensor.

Objective measurement of gaze pattern and eye movement during untethered activity has important applications for neuroscience research and neurological disease detection. Current commercial eye-tracking tools rely on desk-top devices with infrared emitters and conventional frame-based cameras. Although wearable options do exist, the large power-consumption from their conventional cameras limit true long-term mobile usage. The query-driven Dynamic Vision Sensor (qDVS) is a neuromorphic camera which dramatically reduces power consumption by outputting only intensity-change threshold events, as opposed to full frames of intensity data. However, such hardware has not yet been implemented for on-body eye-tracking, but the feasibility can be demonstrated using a mathematical simulator to evaluate the eye-tracking ca-pabilities of the qDVS under controlled conditions. Specifically, a framework utilizing a realistic human eye model in the 3D graphics engine, Unity, is presented to enable the controlled and direct comparison of image-based gaze tracking methods. Eye-tracking based on qDVS frames was compared against two different conventional frame eye-tracking methods - the traditional ellipse pupil-fitting algorithm and a deep learning neural network inference model. Gaze accuracy from qDVS frames achieved an average of 93.2% for movement along the primary horizontal axis (pitch angle) and 93.1 % for movement along the primary vertical axis (yaw angle) under 4 different illumination conditions, demonstrating the feasibility for using qDVS hardware cameras for such applications. The quantitative framework for the direct comparison of eye tracking algorithms presented here is made open-source and can be extended to include other eye parameters, such as pupil dilation, reflection, motion artifact, and more.

Photon counting, censor corrections, and lifetime imaging for improved detection in two-photon microscopy.

We present a high-speed photon counter for use with two-photon microscopy. Counting pulses of photocurrent, as opposed to analog integration, maximizes the signal-to-noise ratio so long as the uncertainty in the count does not exceed the gain-noise of the photodetector. Our system extends this improvement through an estimate of the count that corrects for the censored period after detection of an emission event. The same system can be rapidly reconfigured in software for fluorescence lifetime imaging, which we illustrate by distinguishing between two spectrally similar fluorophores in an in vivo model of microstroke.