Hardware-aware training for large-scale and diverse deep learning inference workloads using in-memory computing-based accelerators.

Analog in-memory computing-a promising approach for energy-efficient acceleration of deep learning workloads-computes matrix-vector multiplications but only approximately, due to nonidealities that often are non-deterministic or nonlinear. This can adversely impact the achievable inference accuracy. Here, we develop an hardware-aware retraining approach to systematically examine the accuracy of analog in-memory computing across multiple network topologies, and investigate sensitivity and robustness to a broad set of nonidealities. By introducing a realistic crossbar model, we improve significantly on earlier retraining approaches. We show that many larger-scale deep neural networks-including convnets, recurrent networks, and transformers-can in fact be successfully retrained to show iso-accuracy with the floating point implementation. Our results further suggest that nonidealities that add noise to the inputs or outputs, not the weights, have the largest impact on accuracy, and that recurrent networks are particularly robust to all nonidealities.

Integrated biosensor platform based on graphene transistor arrays for real-time high-accuracy ion sensing.

Two-dimensional materials such as graphene have shown great promise as biosensors, but suffer from large device-to-device variation due to non-uniform material synthesis and device fabrication technologies. Here, we develop a robust bioelectronic sensing platform  composed of  more than 200 integrated sensing units, custom-built high-speed readout electronics, and machine learning inference that overcomes these challenges to achieve rapid, portable, and reliable measurements. The platform demonstrates reconfigurable multi-ion electrolyte sensing capability and provides highly sensitive, reversible, and real-time response for potassium, sodium, and calcium ions in complex solutions despite variations in device performance. A calibration method leveraging the sensor redundancy and device-to-device variation is also proposed, while a machine learning model trained with multi-dimensional information collected through the multiplexed sensor array is used to enhance the sensing system's functionality and accuracy in ion classification.

Optimised weight programming for analogue memory-based deep neural networks.

Analogue memory-based deep neural networks provide energy-efficiency and per-area throughput gains relative to state-of-the-art digital counterparts such as graphics processing units. Recent advances focus largely on hardware-aware algorithmic training and improvements to circuits, architectures, and memory devices. Optimal translation of software-trained weights into analogue hardware weights-given the plethora of complex memory non-idealities-represents an equally important task. We report a generalised computational framework that automates the crafting of complex weight programming strategies to minimise accuracy degradations during inference, particularly over time. The framework is agnostic to network structure and generalises well across recurrent, convolutional, and transformer neural networks. As a highly flexible numerical heuristic, the approach accommodates arbitrary device-level complexity, making it potentially relevant for a variety of analogue memories. By quantifying the limit of achievable inference accuracy, it also enables analogue memory-based deep neural network accelerators to reach their full inference potential.

Toward Software-Equivalent Accuracy on Transformer-Based Deep Neural Networks With Analog Memory Devices.

Recent advances in deep learning have been driven by ever-increasing model sizes, with networks growing to millions or even billions of parameters. Such enormous models call for fast and energy-efficient hardware accelerators. We study the potential of Analog AI accelerators based on Non-Volatile Memory, in particular Phase Change Memory (PCM), for software-equivalent accurate inference of natural language processing applications. We demonstrate a path to software-equivalent accuracy for the GLUE benchmark on BERT (Bidirectional Encoder Representations from Transformers), by combining noise-aware training to combat inherent PCM drift and noise sources, together with reduced-precision digital attention-block computation down to INT6.

Intravenous Amiodarone and Sotalol Impair Contractility and Cardiac Output, but Procainamide Does Not: A Langendorff Study.

INTRODUCTION: Direct comparison of the effects of antiarrhythmic agents on myocardial performance may be useful in choosing between medications in critically ill patients. Studies directly comparing multiple antiarrhythmic medications are lacking. The use of an experimental heart preparation permits examination of myocardial performance under constant loading conditions. METHODS: Hearts of Sprague Dawley rats (n = 35, 402-507 g) were explanted and cannulated in working heart model with fixed preload and afterload. Each heart was then exposed to a 3-hour infusion of procainamide (20 µg/kg/min), esmolol (100 or 200 µg/kg/min), amiodarone (10 or 20 mg/kg/d), sotalol (80 mg/m2/d), or placebo infusions (n = 5 per dose). Cardiac output, contractility (dP/dTmax), diastolic performance (dP/dTmin), and heart rate were compared between groups over time by linear mixed modeling. RESULTS: Compared with placebo, sotalol decreased contractility by an average of 24% ( P < .001) over the infusion period, as did amiodarone (low dose by 13%, P = .029; high dose by 14%, P = .013). Compared with placebo, mean cardiac output was significantly lower in animals treated with sotalol (by 22%, P = .016) and esmolol 200 µg/kg/min (by 23%, P = .012). Over time, amiodarone decreased cardiac output (20 mg/kg/d, ß = -89 [-144, -33] µL/min2 decrease, P = .002) and also worsened diastolic function, decreasing dP/dTmin by ∼18% and 22% ( P = .032 and P = .011, low and high doses, respectively). Procainamide did not have a significant effect on any measures of systolic or diastolic performance. CONCLUSIONS: In isolated hearts, amiodarone and sotalol depressed myocardial contractility, cardiac output, and diastolic function. However, procainamide did not negatively affect myocardial performance and represents a favorable agent in settings of therapeutic equivalence.

Chemiresistive Graphene Sensors for Ammonia Detection.

The primary objective of this work is to demonstrate a novel sensor system as a convenient vehicle for scaled-up repeatability and the kinetic analysis of a pixelated testbed. This work presents a sensor system capable of measuring hundreds of functionalized graphene sensors in a rapid and convenient fashion. The sensor system makes use of a novel array architecture requiring only one sensor per pixel and no selector transistor. The sensor system is employed specifically for the evaluation of Co(tpfpp)ClO4 functionalization of graphene sensors for the detection of ammonia as an extension of previous work. Co(tpfpp)ClO4 treated graphene sensors were found to provide 4-fold increased ammonia sensitivity over pristine graphene sensors. Sensors were also found to exhibit excellent selectivity over interfering compounds such as water and common organic solvents. The ability to monitor a large sensor array with 160 pixels provides insights into performance variations and reproducibility-critical factors in the development of practical sensor systems. All sensors exhibit the same linearly related responses with variations in response exhibiting Gaussian distributions, a key finding for variation modeling and quality engineering purposes. The mean correlation coefficient between sensor responses was found to be 0.999 indicating highly consistent sensor responses and excellent reproducibility of Co(tpfpp)ClO4 functionalization. A detailed kinetic model is developed to describe sensor response profiles. The model consists of two adsorption mechanisms-one reversible and one irreversible-and is shown capable of fitting experimental data with a mean percent error of 0.01%.

Frequency Response of Graphene Electrolyte-Gated Field-Effect Transistors.

This work develops the first frequency-dependent small-signal model for graphene electrolyte-gated field-effect transistors (EGFETs). Graphene EGFETs are microfabricated to measure intrinsic voltage gain, frequency response, and to develop a frequency-dependent small-signal model. The transfer function of the graphene EGFET small-signal model is found to contain a unique pole due to a resistive element, which stems from electrolyte gating. Intrinsic voltage gain, cutoff frequency, and transition frequency for the microfabricated graphene EGFETs are approximately 3.1 V/V, 1.9 kHz, and 6.9 kHz, respectively. This work marks a critical step in the development of high-speed chemical and biological sensors using graphene EGFETs.

Correction: Large-scale sensor systems based on graphene electrolyte-gated field-effect transistors.

Correction for 'Large-scale sensor systems based on graphene electrolyte-gated field-effect transistors' by Charles Mackin, et al., Analyst, 2016, 141, 2704-2711.

Large-scale sensor systems based on graphene electrolyte-gated field-effect transistors.

This work reports a novel graphene electrolyte-gated field-effect transistor (EGFET) array architecture along with a compact, self-contained, and inexpensive measurement system that allows DC characterization of hundreds of graphene EGFETs as a function of VDS and VGS within a matter of minutes. We develop a reliable graphene EGFET fabrication process capable of producing 100% yield for a sample size of 256 devices. Large sample size statistical analysis of graphene EGFET electrical performance is performed for the first time. This work develops a compact piecewise DC model for graphene EGFETs that is shown capable of fitting 87% of IDSvs. VGS curves with a mean percent error of 7% or less. The model is used to extract variations in device parameters such as mobility, contact resistance, minimum carrier concentration, and Dirac point. Correlations in variations are presented. Lastly, this work presents a framework for application-specific optimization of large-scale sensor designs based on graphene EGFETs.

Structural Distortion of Molybdenum-Doped Manganese Oxide Octahedral Molecular Sieves for Enhanced Catalytic Performance.

Due to the excellent catalytic performance of manganese oxide (K-OMS-2) in a wide range of applications, incorporation of various dopants has been commonly applied for K-OMS-2 to acquire additional functionality or activities. However, the understanding of its substitution mechanism with respect to the catalytic performance of doped K-OMS-2 materials remains unclear. Here we present the structural distortion (from tetragonal to monoclinic cell) and morphological evolution in K-OMS-2 materials by doping hexavalent molybdenum. With a Mo-to-Mn ratio of 1:20 (R-1:20) in the preparation, the resultant monoclinic K-OMS-2 shows a small equidimensional particle size (∼15 nm), a high surface area of 213 m(2) g(-1), and greatly improved catalytic activity toward CO oxidation with lower onset temperatures (40 °C) than that of pristine K-OMS-2 (above 130 °C). HR-TEM analyses reveal direct evidence of structural distortion on the cross-section of 2 × 2 tunnels with the absence of 4-fold rotation symmetry expected for a tetragonal cell, which are indexed using a monoclinic cell. Our results suggest that substitution of Mo(6+) for Mn(3+) (rather than Mn(4+)) coupled with the vacancy generation results in a distorted structure and unique morphology. The weakened Mn-O bonds and Mn vacancies associated with the structural distortion may be mainly responsible for the enhanced catalytic activity of monoclinic K-OMS-2 instead of dopant species.

Role of interfacial oxide in high-efficiency graphene-silicon Schottky barrier solar cells.

The advent of chemical vapor deposition (CVD) grown graphene has allowed researchers to investigate large area graphene/n-silicon Schottky barrier solar cells. Using chemically doped graphene, efficiencies of nearly 10% can be achieved for devices without antireflective coatings. However, many devices reported in past literature often exhibit a distinctive s-shaped kink in the measured I/V curves under illumination resulting in poor fill factor. This behavior is especially prevalent for devices with pristine (not chemically doped) graphene but can be seen in some cases for doped graphene as well. In this work, we show that the native oxide on the silicon presents a transport barrier for photogenerated holes and causes recombination current, which is responsible for causing the kink. We experimentally verify our hypothesis and propose a simple semiconductor physics model that qualitatively captures the effect. Furthermore, we offer an additional optimization to graphene/n-silicon devices: by choosing the optimal oxide thickness, we can increase the efficiency of our devices to 12.4% after chemical doping and to a new record of 15.6% after applying an antireflective coating.

Elevated double negative T cells in pediatric autoimmunity.

PURPOSE: Autoimmune diseases are thought to be caused by a loss of self-tolerance of the immune system. One candidate marker of immune dysregulation in autoimmune disease is the presence of increased double negative T cells (DNTs) in the periphery. DNTs are characteristically elevated in autoimmune lymphoproliferative syndrome, a systemic autoimmune disease caused by defective lymphocyte apoptosis due to Fas pathway defects. DNTs have also been found in the peripheral blood of adult patients with systemic lupus erythematosus (SLE), where they may be pathogenic. DNTs in children with autoimmune disease have not been investigated. METHODS: We evaluated DNTs in pediatric patients with SLE, mixed connective tissue disease (MCTD), juvenile idiopathic arthritis (JIA), or elevated antinuclear antibody (ANA) but no systemic disease. DNTs (CD3(+)CD56(-)TCRαß(+)CD4(-)CD8(-)) from peripheral blood mononuclear cells were analyzed by flow cytometry from 54 pediatric patients including: 23 SLE, 15 JIA, 11 ANA and 5 MCTD compared to 28 healthy controls. RESULTS: Sixteen cases (29.6 %) had elevated DNTs (≥2.5 % of CD3(+)CD56(-)TCRαß(+) cells) compared to 1 (3.6 %) control. Medication usage including cytotoxic drugs and absolute lymphocyte count were not associated with DNT levels, and percentages of DNTs were stable over time. Analysis of multiple phenotypic and activation markers showed increased CD45RA expression on DNTs from patients with autoimmune disease compared to controls. CONCLUSION: DNTs are elevated in a subset of pediatric patients with autoimmune disease and additional investigations are required to determine their precise role in autoimmunity.