Impact of white noise in artificial neural networks trained for classification: Performance and noise mitigation strategies.

In recent years, the hardware implementation of neural networks, leveraging physical coupling and analog neurons has substantially increased in relevance. Such nonlinear and complex physical networks provide significant advantages in speed and energy efficiency, but are potentially susceptible to internal noise when compared to digital emulations of such networks. In this work, we consider how additive and multiplicative Gaussian white noise on the neuronal level can affect the accuracy of the network when applied for specific tasks and including a softmax function in the readout layer. We adapt several noise reduction techniques to the essential setting of classification tasks, which represent a large fraction of neural network computing. We find that these adjusted concepts are highly effective in mitigating the detrimental impact of noise.

Noise-mitigation strategies in physical feedforward neural networks.

Physical neural networks are promising candidates for next generation artificial intelligence hardware. In such architectures, neurons and connections are physically realized and do not leverage digital concepts with their practically infinite signal-to-noise ratio to encode, transduce, and transform information. They, therefore, are prone to noise with a variety of statistical and architectural properties, and effective strategies leveraging network-inherent assets to mitigate noise in a hardware-efficient manner are important in the pursuit of next generation neural network hardware. Based on analytical derivations, we here introduce and analyze a variety of different noise-mitigation approaches. We analytically show that intra-layer connections in which the connection matrix's squared mean exceeds the mean of its square fully suppress uncorrelated noise. We go beyond and develop two synergistic strategies for noise that is uncorrelated and correlated across populations of neurons. First, we introduce the concept of ghost neurons, where each group of neurons perturbed by correlated noise has a negative connection to a single neuron, yet without receiving any input information. Second, we show that pooling of neuron populations is an efficient approach to suppress uncorrelated noise. As such, we developed a general noise-mitigation strategy leveraging the statistical properties of the different noise terms most relevant in analog hardware. Finally, we demonstrate the effectiveness of this combined approach for a trained neural network classifying the modified National Institute of Standards and Technology handwritten digits, for which we achieve a fourfold improvement of the output signal-to-noise ratio. Our noise mitigation lifts the 92.07% classification accuracy of the noisy neural network to 97.49%, which is essentially identical to the 97.54% of the noise-free network.

Effects of extreme meteorological conditions in 2018 on European methane emissions estimated using atmospheric inversions.

The effect of the 2018 extreme meteorological conditions in Europe on methane (CH4) emissions is examined using estimates from four atmospheric inversions calculated for the period 2005-2018. For most of Europe, we find no anomaly in 2018 compared to the 2005-2018 mean. However, we find a positive anomaly for the Netherlands in April, which coincided with positive temperature and soil moisture anomalies suggesting an increase in biogenic sources. We also find a negative anomaly for the Netherlands for September-October, which coincided with a negative anomaly in soil moisture, suggesting a decrease in soil sources. In addition, we find a positive anomaly for Serbia in spring, summer and autumn, which coincided with increases in temperature and soil moisture, again suggestive of changes in biogenic sources, and the annual emission for 2018 was 33 ± 38% higher than the 2005-2017 mean. These results indicate that CH4 emissions from areas where the natural source is thought to be relatively small can still vary due to meteorological conditions. At the European scale though, the degree of variability over 2005-2018 was small, and there was negligible impact on the annual CH4 emissions in 2018 despite the extreme meteorological conditions. This article is part of a discussion meeting issue 'Rising methane: is warming feeding warming? (part 2)'.

Optical pumping of quantum dot micropillar lasers.

Arrays of quantum dot micropillar lasers are an attractive technology platform for various applications in the wider field of nanophotonics. Of particular interest is the potential efficiency enhancement as a consequence of cavity quantum electrodynamics effects, which makes them prime candidates for next generation photonic neurons in neural network hardware. However, particularly for optical pumping, their power-conversion efficiency can be very low. Here we perform an in-depth experimental analysis of quantum dot microlasers and investigate their input-output relationship over a wide range of optical pumping conditions. We find that the current energy efficiency limitation is caused by disadvantageous optical pumping concepts and by a low exciton conversion efficiency. Our results indicate that for non-resonant pumping into the GaAs matrix (wetting layer), 3.4% (0.6%) of the optical pump is converted into lasing-relevant excitons, and of those only 2% (0.75%) provide gain to the lasing transition. Based on our findings, we propose to improve the pumping efficiency by orders of magnitude by increasing the aluminium content of the AlGaAs/GaAs mirror pairs in the upper Bragg reflector.

The fingerprint of the 2018 drought in Europe on ground-based atmospheric CO2 measurements.

During the summer of 2018, a widespread drought developed over Northern and Central Europe. The increase in temperature and the reduction of soil moisture have influenced carbon dioxide (CO2) exchange between the atmosphere and terrestrial ecosystems in various ways, such as a reduction of photosynthesis, changes in ecosystem respiration, or allowing more frequent fires. In this study, we characterize the resulting perturbation of the atmospheric CO2 seasonal cycles. 2018 has a good coverage of European regions affected by drought, allowing the investigation of how ecosystem flux anomalies impacted spatial CO2 gradients between stations. This density of stations is unprecedented compared to previous drought events in 2003 and 2015, particularly thanks to the deployment of the Integrated Carbon Observation System (ICOS) network of atmospheric greenhouse gas monitoring stations in recent years. Seasonal CO2 cycles from 48 European stations were available for 2017 and 2018. Earlier data were retrieved for comparison from international databases or national networks. Here, we show that the usual summer minimum in CO2 due to the surface carbon uptake was reduced by 1.4 ppm in 2018 for the 10 stations located in the area most affected by the temperature anomaly, mostly in Northern Europe. Notwithstanding, the CO2 transition phases before and after July were slower in 2018 compared to 2017, suggesting an extension of the growing season, with either continued CO2 uptake by photosynthesis and/or a reduction in respiration driven by the depletion of substrate for respiration inherited from the previous months due to the drought. For stations with sufficiently long time series, the CO2 anomaly observed in 2018 was compared to previous European droughts in 2003 and 2015. Considering the areas most affected by the temperature anomalies, we found a higher CO2 anomaly in 2003 (+3 ppm averaged over 4 sites), and a smaller anomaly in 2015 (+1 ppm averaged over 11 sites) compared to 2018. This article is part of the theme issue 'Impacts of the 2018 severe drought and heatwave in Europe: from site to continental scale'.

Fundamental aspects of noise in analog-hardware neural networks.

We study and analyze the fundamental aspects of noise propagation in recurrent as well as deep, multilayer networks. The motivation of our study is neural networks in analog hardware; yet, the methodology provides insight into networks in general. Considering noisy linear nodes, we investigate the signal-to-noise ratio at the network's outputs, which determines the upper limit of computational precision. We consider additive and multiplicative noise, which can be purely local as well as correlated across populations of neurons. This covers the chief internal-perturbations of hardware networks, and noise amplitudes were obtained from a physically implemented neural network. Analytically derived descriptions agree exceptionally well with numerical data, enabling clear identification of the components critical for management and mitigation of noise. We find that analog neural networks are surprisingly robust, in particular, against noisy neurons. Their uncorrelated perturbations are almost fully suppressed, while correlated noise can accumulate. Our work identifies notoriously sensitive points while highlighting a surprising robustness of such computational systems.

Outlier classification using autoencoders: Application for fluctuation driven flows in fusion plasmas.

Understanding the statistics of fluctuation driven flows in the boundary layer of magnetically confined plasmas is desired to accurately model the lifetime of the vacuum vessel components. Mirror Langmuir probes (MLPs) are a novel diagnostic that uniquely allow us to sample the plasma parameters on a time scale shorter than the characteristic time scale of their fluctuations. Sudden large-amplitude fluctuations in the plasma degrade the precision and accuracy of the plasma parameters reported by MLPs for cases in which the probe bias range is of insufficient amplitude. While some data samples can readily be classified as valid and invalid, we find that such a classification may be ambiguous for up to 40% of data sampled for the plasma parameters and bias voltages considered in this study. In this contribution, we employ an autoencoder (AE) to learn a low-dimensional representation of valid data samples. By definition, the coordinates in this space are the features that mostly characterize valid data. Ambiguous data samples are classified in this space using standard classifiers for vectorial data. In this way, we avoid defining complicated threshold rules to identify outliers, which require strong assumptions and introduce biases in the analysis. By removing the outliers that are identified in the latent low-dimensional space of the AE, we find that the average conductive and convective radial heat fluxes are between approximately 5% and 15% lower as when removing outliers identified by threshold values. For contributions to the radial heat flux due to triple correlations, the difference is up to 40%.

Two-dimensional spatiotemporal complexity in dual-delayed nonlinear feedback systems: Chimeras and dissipative solitons.

We demonstrate for a photonic nonlinear system that two highly asymmetric feedback delays can induce a variety of emergent patterns which are highly robust during the system's global evolution. Explicitly, two-dimensional chimeras and dissipative solitons become visible upon a space-time transformation. Switching between chimeras and dissipative solitons requires only adjusting two system parameters, demonstrating self-organization exclusively based on the system's dynamical properties. Experiments were performed using a tunable semiconductor laser's transmission through a Fabry-Pérot resonator resulting in an Airy function as nonlinearity. Resulting dynamics were bandpass filtered and propagated along two feedback paths whose time delays differ by two orders of magnitude. An excellent agreement between experimental results and the theoretical model given by modified Ikeda equations was achieved.

Key Issues for Seamless Integrated Chemistry-Meteorology Modeling.

Online coupled meteorology-atmospheric chemistry models have greatly evolved in recent years. Although mainly developed by the air quality modeling community, these integrated models are also of interest for numerical weather prediction and climate modeling, as they can consider both the effects of meteorology on air quality and the potentially important effects of atmospheric composition on weather. This paper summarizes the main conclusions from the "Symposium on Coupled Chemistry-Meteorology/Climate Modelling: Status and Relevance for Numerical Weather Prediction, Air Quality and Climate Research," which was initiated by the European COST Action ES1004 "European Framework for Online Integrated Air Quality and Meteorology Modelling (EuMetChem)." It offers a brief review of the current status of online coupled meteorology and atmospheric chemistry modeling and a survey of processes relevant to the interactions between atmospheric physics, dynamics, and composition. In addition, it highlights scientific issues and emerging challenges that require proper consideration to improve the reliability and usability of these models for three main application areas: air quality, meteorology (including weather prediction), and climate modeling. It presents a synthesis of scientific progress in the form of answers to nine key questions, and provides recommendations for future research directions and priorities in the development, application, and evaluation of online coupled models.

Observation of Efficient Lower Hybrid Current Drive at High Density in Diverted Plasmas on the Alcator C-Mod Tokamak.

Efficient lower hybrid current drive (LHCD) is demonstrated at densities up to n[over ¯]_{e}≈1.5×10^{20} m^{-3} in diverted plasmas on the Alcator C-Mod tokamak by operating at increased plasma current-and therefore reduced Greenwald density fraction. This density exceeds the nominal "LH density limit" at n[over ¯]_{e}≈1.0×10^{20} m^{-3} reported previously, above which an anomalous loss of current drive efficiency was observed. The recovery of current drive efficiency to a level consistent with engineering scalings is correlated with a reduction in density shoulders and turbulence levels in the far scrape-off layer. Concurrently, rf wave interaction with the edge and/or scrape-off-layer plasma is reduced, as indicated by a minimal broadening of the wave frequency spectrum measured at the plasma edge. These results have important implications for sustaining steady-state tokamak operation and indicate a pathway forward for implementing efficient LHCD in a reactor.

Magnetic tweezers optimized to exert high forces over extended distances from the magnet in multicellular systems.

Magnetic tweezers are mainly divided into two classes depending on the ability of applying torque or forces to the magnetic probe. We focused on the second category and designed a device composed by a single electromagnet equipped with a core having a special asymmetric profile to exert forces as large as 230 pN-2.8 µm Dynabeads at distances in excess of 100 µm from the magnetic tip. Compared to existing solutions our magnetic tweezers overcome important limitations, opening new experimental paths for the study of a wide range of materials in a variety of biophysical research settings. We discuss the benefits and drawbacks of different magnet core characteristics, which led us to design the current core profile. To demonstrate the usefulness of our magnetic tweezers, we determined the microrheological properties inside embryos of Drosophila melanogaster during the syncytial stage. Measurements in different locations along the dorsal-ventral axis of the embryos showed little variation, with a slight increase in cytoplasm viscosity at the periphery of the embryos. The mean cytoplasm viscosity we obtain by active force exertion inside the embryos is comparable to that determined passively using high-speed video microrheology.

The flush-mounted rail Langmuir probe array designed for the Alcator C-Mod vertical target plate divertor.

An array of flush-mounted and toroidally elongated Langmuir probes (henceforth called rail probes) have been specifically designed for the Alcator C-Mod's vertical target plate divertor and operated over multiple campaigns. The "flush" geometry enables the tungsten electrodes to survive high heat flux conditions in which traditional "proud" tungsten electrodes suffer damage from melting. The toroidally elongated rail-like geometry reduces the influence of sheath expansion, which is an important effect to consider in the design and interpretation of flush-mounted Langmuir probes. The new rail probes successfully operated during C-Mod's FY2015 and FY2016 experimental campaigns with no evidence of damage, despite being regularly subjected to heat flux densities parallel to the magnetic field exceeding ∼1 GW m-2 for short periods of time. A comparison between rail and proud probe data indicates that sheath expansion effects were successfully mitigated by the rail design, extending the use of these Langmuir probes to incident magnetic field line angles as low as 0.5°.

Comprehensive analysis of two Shank3 and the Cacna1c mouse models of autism spectrum disorder.

To expand, analyze and extend published behavioral phenotypes relevant to autism spectrum disorder (ASD), we present a study of three ASD genetic mouse models: Feng's Shank3tm2Gfng model, hereafter Shank3/F, Jiang's Shank3tm1Yhj model, hereafter Shank3/J and the Cacna1c deletion model. The Shank3 models mimick gene mutations associated with Phelan-McDermid Syndrome and the Cacna1c model recapitulates the deletion underlying Timothy syndrome. This study utilizes both standard and novel behavioral tests with the same methodology used in our previously published companion report on the Cntnap2 null and 16p11.2 deletion models. We found that some but not all behaviors replicated published findings and those that did replicate, such as social behavior and overgrooming in Shank3 models, tended to be milder than reported elsewhere. The Shank3/F model, and to a much lesser extent, the Shank3/J and Cacna1c models, showed hypoactivity and a general anxiety-like behavior triggered by external stimuli which pervaded social interactions. We did not detect deficits in a cognitive procedural learning test nor did we observe perseverative behavior in these models. We did, however, find differences in exploratory patterns of Cacna1c mutant mice suggestive of a behavioral effect in a social setting. In addition, only Shank3/F showed differences in sensory-gating. Both positive and negative results from this study will be useful in identifying the most robust and replicable behavioral signatures within and across mouse models of autism. Understanding these phenotypes may shed light of which features to study when screening compounds for potential therapeutic interventions.

Role of dynamical injection locking and characteristic pulse events for low frequency fluctuations in semiconductor lasers.

We investigate the dynamics of semiconductor lasers subject to time-delayed optical feedback from the perspective of dynamical self-injection locking. Based on the Lang-Kobayashi model, we perform an analysis of the well-known Low Frequency Fluctuations (LFFs) in the frequency-intensity plane. Moreover, we investigate a recently found dynamical regime of fragmented LFFs by means of a locking-range analysis, spectral comparison and precursor pulse identification. We show that LFF dynamics can be explained by dynamical optical injection locking due to the delayed optical feedback. Moreover, the fragmented LFFs occur due to a re-injection locking induced by a particular optical pulse structure in the chaotic feedback dynamics. This is corroborated by experiments with a semiconductor laser experiencing delayed feedback from an optical fiber loop. The dynamical nature of the feedback injection results in an eventual loss, but also possible regaining, of the locking, explaining the recently observed phenomenon of fragmented LFFs.

SEN1500, a novel oral amyloid-ß aggregation inhibitor, attenuates brain pathology in a mouse model of Alzheimer's disease.

INTRODUCTION: Amyloid-ß (Aß) aggregation is thought to be a major pathogenic event underlying the neuropathology of Alzheimer's disease (AD). The development of new drugs inhibiting the Aß aggregation process is, therefore, important. SEN1500, an orally bioavailable and CNS-penetrant Aß aggregation inhibitor, has previously been shown to reduce spatial learning and memory deficits in an APP transgenic mouse model. To verify that the pharmacological properties of SEN1500 are not unique to this model, we investigated brain Aß pathology, neuroinflammation, as well as memory in a different mouse model of AD expressing the human amyloid precursor protein with Swedish and London mutations (APPSL). MATERIALS & METHODS: APPSL transgenic mice and non-transgenic littermates were treated with SEN1500 via food pellets from three months of age for four months. At the end of the treatment, animals were tested for memory deficits using the contextual fear conditioning test and brain tissue was analyzed for soluble and insoluble amyloid-ß1-38, -40, -42, ß-amyloid plaques, ß-sheet plaque cores, as well as for astrocytosis and activated microglia. RESULTS: SEN1500 treatment lowered insoluble Aß levels and ß-amyloid plaque load in the brain compared with control-treated APPSL mice. Activated microglia were significantly reduced in the cortex but not the hippocampus of SEN1500-treated APPSL mice. Memory deficits of APPSL mice could not be rescued by SEN1500. DISCUSSION: SEN1500 is not only able to reduce Aß pathology and activated microglia but also to improve learning and memory as previously shown, making SEN1500 a potential candidate for human AD treatment. This Aß aggregation inhibitor could be a promising therapeutic agent for the disease-modifying treatment of AD.

Effects of dynamic-demand-control appliances on the power grid frequency.

Power grid frequency control is a demanding task requiring expensive idle power plants to adapt the supply to the fluctuating demand. An alternative approach is controlling the demand side in such a way that certain appliances modify their operation to adapt to the power availability. This is especially important to achieve a high penetration of renewable energy sources. A number of methods to manage the demand side have been proposed. In this work we focus on dynamic demand control (DDC), where smart appliances can delay their switchings depending on the frequency of the system. We introduce a simple model to study the effects of DDC on the frequency of the power grid. The model includes the power plant equations, a stochastic model for the demand that reproduces, adjusting a single parameter, the statistical properties of frequency fluctuations measured experimentally, and a generic DDC protocol. We find that DDC can reduce small and medium-size fluctuations but it can also increase the probability of observing large frequency peaks due to the necessity of recovering pending task. We also conclude that a deployment of DDC around 30-40% already allows a significant reduction of the fluctuations while keeping the number of pending tasks low.

Linear servomotor probe drive system with real-time self-adaptive position control for the Alcator C-Mod tokamak.

A new servomotor drive system has been developed for the horizontal reciprocating probe on the Alcator C-Mod tokamak. Real-time measurements of plasma temperature and density-through use of a mirror Langmuir probe bias system-combined with a commercial linear servomotor and controller enable self-adaptive position control. Probe surface temperature and its rate of change are computed in real time and used to control probe insertion depth. It is found that a universal trigger threshold can be defined in terms of these two parameters; if the probe is triggered to retract when crossing the trigger threshold, it will reach the same ultimate surface temperature, independent of velocity, acceleration, or scrape-off layer heat flux scale length. In addition to controlling the probe motion, the controller is used to monitor and control all aspects of the integrated probe drive system.

Semiconductor laser linewidth reduction by six orders of magnitude via delayed optical feedback.

We demonstrate a coherence increase by six orders of magnitude of a standard quantum well semiconductor laser. Using a simple, optical-fiber-based feedback scheme, we stabilize the laser in a high-gain mode of a long external cavity. In a modified self-heterodyne measurement, we mix the high-gain mode with a strongly suppressed side mode and obtain an interference linewidth of only 12.6 Hz, corresponding to a decoherence of (3.1±2.9) Hz. In an independent characterization using an etalon, we deduce an upper limit of 300 Hz for the laser linewidth. The laser stably resides in this mode for tens of seconds. Our results agree with theoretical predictions.

Feedback system for divertor impurity seeding based on real-time measurements of surface heat flux in the Alcator C-Mod tokamak.

Mitigation of the intense heat flux to the divertor is one of the outstanding problems in fusion energy. One technique that has shown promise is impurity seeding, i.e., the injection of low-Z gaseous impurities (typically N2 or Ne) to radiate and dissipate the power before it arrives to the divertor target plate. To this end, the Alcator C-Mod team has created a first-of-its-kind feedback system to control the injection of seed gas based on real-time surface heat flux measurements. Surface thermocouples provide real-time measurements of the surface temperature response to the plasma heat flux. The surface temperature measurements are inputted into an analog computer that "solves" the 1-D heat transport equation to deliver accurate, real-time signals of the surface heat flux. The surface heat flux signals are sent to the C-Mod digital plasma control system, which uses a proportional-integral-derivative (PID) algorithm to control the duty cycle demand to a pulse width modulated piezo valve, which in turn controls the injection of gas into the private flux region of the C-Mod divertor. This paper presents the design and implementation of this new feedback system as well as initial results using it to control divertor heat flux.

A Unified Framework for Reservoir Computing and Extreme Learning Machines based on a Single Time-delayed Neuron.

In this paper we present a unified framework for extreme learning machines and reservoir computing (echo state networks), which can be physically implemented using a single nonlinear neuron subject to delayed feedback. The reservoir is built within the delay-line, employing a number of "virtual" neurons. These virtual neurons receive random projections from the input layer containing the information to be processed. One key advantage of this approach is that it can be implemented efficiently in hardware. We show that the reservoir computing implementation, in this case optoelectronic, is also capable to realize extreme learning machines, demonstrating the unified framework for both schemes in software as well as in hardware.