skip to main content


Title: Low-loss, fine resolution wavelength-selective switch realized at a silicon photonics foundry

We report a demonstration of a 3-channel wavelength-selective switch with individual channel bandwidths of 2 GHz and drop port loss below 1 dB, paving the way for efficient spectrum utilization in quantum networking applications.

 
more » « less
Award ID(s):
1747426
NSF-PAR ID:
10479854
Author(s) / Creator(s):
; ; ; ; ; ; ;
Publisher / Repository:
Optica Publishing Group
Date Published:
Page Range / eLocation ID:
STu3J.6
Format(s):
Medium: X
Location:
San Jose, CA
Sponsoring Org:
National Science Foundation
More Like this
  1. Abstract

    Despite its relatively small magnitude, cross-channel circulation in estuaries can influence the along-channel momentum balance, dispersion, and transport. We investigate spatial and temporal variation in cross-channel circulation at two contrasting sites in the Hudson River estuary. The two sites differ in the relative strength and direction of Coriolis and curvature forcing. We contrast the patterns and magnitudes of flow at the two sites during varying conditions in stratification driven by tidal amplitude and river discharge. We found well-defined flows during flood tides at both sites, characterized by mainly two-layer structures when the water column was more homogeneous and structures with three or more layers when the water column was more stratified. Ebb tides had generally weaker and less definite flows, except at one site where curvature and Coriolis reinforced each other during spring tide ebbs. Cross-channel currents had similar patterns, but were oppositely directed at the two sites, demonstrating the importance of curvature even in channels with relatively gradual curves. Coriolis and curvature dominated the measured terms in the cross-channel momentum balance. Their combination was generally consistent with driving the observed patterns and directions of flow, but local acceleration and cross-channel advection made some notable contributions. A large residual in the momentum balance indicates that some combination of vertical stress divergence, baroclinic pressure gradients, and along-channel and vertical advection must play an essential role, but data limitations prevented an accurate estimation of these terms. Cross-channel advection affected the along-channel momentum balance at times, with implications for the exchange flow’s strength.

    Significance Statement

    Currents that flow across the channel in an estuary move slower than those flowing along the channel, but they can transport materials and change water properties in important ways, affecting human uses of estuaries such as shipping, aquaculture, and recreation. We wanted to better understand cross-channel currents in the Hudson River estuary. We found that larger tides produced the strongest cross-channel currents with a two-layer pattern, compared to weaker currents with three layers during smaller tides. Higher or lower river flow also affected current strength. Comparing two locations, we saw cross-channel currents moving in opposite directions because of differences in the curvature of the river channel. Our results show how channel curvature and Earth’s rotation combine to produce cross-channel currents.

     
    more » « less
  2. Key points

    Gap junctions formed by different connexins are expressed throughout the body and harbour unique channel properties that have not been fully defined mechanistically.

    Recent structural studies by cryo‐electron microscopy have produced high‐resolution models of the related but functionally distinct lens connexins (Cx50 and Cx46) captured in a stable open state, opening the door for structure–function comparison.

    Here, we conducted comparative molecular dynamics simulation and electrophysiology studies to dissect the isoform‐specific differences in Cx46 and Cx50 intercellular channel function.

    We show that key determinants Cx46 and Cx50 gap junction channel open stability and unitary conductance are shaped by structural and dynamic features of their N‐terminal domains, in particular the residue at the 9th position and differences in hydrophobic anchoring sites.

    The results of this study establish the open state Cx46/50 structural models as archetypes for structure–function studies targeted at elucidating the mechanism of gap junction channels and the molecular basis of disease‐causing variants.

    Abstract

    Connexins form intercellular communication channels, known as gap junctions (GJs), that facilitate diverse physiological roles, from long‐range electrical and chemical coupling to coordinating development and nutrient exchange. GJs formed by different connexin isoforms harbour unique channel properties that have not been fully defined mechanistically. Recent structural studies on Cx46 and Cx50 defined a novel and stable open state and implicated the amino‐terminal (NT) domain as a major contributor for isoform‐specific functional differences between these closely related lens connexins. To better understand these differences, we constructed models corresponding to wildtype Cx50 and Cx46 GJs, NT domain swapped chimeras, and point variants at the 9th residue for comparative molecular dynamics (MD) simulation and electrophysiology studies. All constructs formed functional GJ channels, except the chimeric Cx46‐50NT variant, which correlated with an introduced steric clash and increased dynamical behaviour (instability) of the NT domain observed by MD simulation. Single channel conductance correlated well with free‐energy landscapes predicted by MD, but resulted in a surprisingly greater degree of effect. Additionally, we observed significant effects on transjunctional voltage‐dependent gating (Vjgating) and/or open state dwell times induced by the designed NT domain variants. Together, these studies indicate intra‐ and inter‐subunit interactions involving both hydrophobic and charged residues within the NT domains of Cx46 and Cx50 play important roles in defining GJ open state stability and single channel conductance, and establish the open state Cx46/50 structural models as archetypes for structure–function studies targeted at elucidating GJ channel mechanisms and the molecular basis of cataract‐linked connexin variants.

     
    more » « less
  3. Obeid, Iyad Selesnick (Ed.)
    Electroencephalography (EEG) is a popular clinical monitoring tool used for diagnosing brain-related disorders such as epilepsy [1]. As monitoring EEGs in a critical-care setting is an expensive and tedious task, there is a great interest in developing real-time EEG monitoring tools to improve patient care quality and efficiency [2]. However, clinicians require automatic seizure detection tools that provide decisions with at least 75% sensitivity and less than 1 false alarm (FA) per 24 hours [3]. Some commercial tools recently claim to reach such performance levels, including the Olympic Brainz Monitor [4] and Persyst 14 [5]. In this abstract, we describe our efforts to transform a high-performance offline seizure detection system [3] into a low latency real-time or online seizure detection system. An overview of the system is shown in Figure 1. The main difference between an online versus offline system is that an online system should always be causal and has minimum latency which is often defined by domain experts. The offline system, shown in Figure 2, uses two phases of deep learning models with postprocessing [3]. The channel-based long short term memory (LSTM) model (Phase 1 or P1) processes linear frequency cepstral coefficients (LFCC) [6] features from each EEG channel separately. We use the hypotheses generated by the P1 model and create additional features that carry information about the detected events and their confidence. The P2 model uses these additional features and the LFCC features to learn the temporal and spatial aspects of the EEG signals using a hybrid convolutional neural network (CNN) and LSTM model. Finally, Phase 3 aggregates the results from both P1 and P2 before applying a final postprocessing step. The online system implements Phase 1 by taking advantage of the Linux piping mechanism, multithreading techniques, and multi-core processors. To convert Phase 1 into an online system, we divide the system into five major modules: signal preprocessor, feature extractor, event decoder, postprocessor, and visualizer. The system reads 0.1-second frames from each EEG channel and sends them to the feature extractor and the visualizer. The feature extractor generates LFCC features in real time from the streaming EEG signal. Next, the system computes seizure and background probabilities using a channel-based LSTM model and applies a postprocessor to aggregate the detected events across channels. The system then displays the EEG signal and the decisions simultaneously using a visualization module. The online system uses C++, Python, TensorFlow, and PyQtGraph in its implementation. The online system accepts streamed EEG data sampled at 250 Hz as input. The system begins processing the EEG signal by applying a TCP montage [8]. Depending on the type of the montage, the EEG signal can have either 22 or 20 channels. To enable the online operation, we send 0.1-second (25 samples) length frames from each channel of the streamed EEG signal to the feature extractor and the visualizer. Feature extraction is performed sequentially on each channel. The signal preprocessor writes the sample frames into two streams to facilitate these modules. In the first stream, the feature extractor receives the signals using stdin. In parallel, as a second stream, the visualizer shares a user-defined file with the signal preprocessor. This user-defined file holds raw signal information as a buffer for the visualizer. The signal preprocessor writes into the file while the visualizer reads from it. Reading and writing into the same file poses a challenge. The visualizer can start reading while the signal preprocessor is writing into it. To resolve this issue, we utilize a file locking mechanism in the signal preprocessor and visualizer. Each of the processes temporarily locks the file, performs its operation, releases the lock, and tries to obtain the lock after a waiting period. The file locking mechanism ensures that only one process can access the file by prohibiting other processes from reading or writing while one process is modifying the file [9]. The feature extractor uses circular buffers to save 0.3 seconds or 75 samples from each channel for extracting 0.2-second or 50-sample long center-aligned windows. The module generates 8 absolute LFCC features where the zeroth cepstral coefficient is replaced by a temporal domain energy term. For extracting the rest of the features, three pipelines are used. The differential energy feature is calculated in a 0.9-second absolute feature window with a frame size of 0.1 seconds. The difference between the maximum and minimum temporal energy terms is calculated in this range. Then, the first derivative or the delta features are calculated using another 0.9-second window. Finally, the second derivative or delta-delta features are calculated using a 0.3-second window [6]. The differential energy for the delta-delta features is not included. In total, we extract 26 features from the raw sample windows which add 1.1 seconds of delay to the system. We used the Temple University Hospital Seizure Database (TUSZ) v1.2.1 for developing the online system [10]. The statistics for this dataset are shown in Table 1. A channel-based LSTM model was trained using the features derived from the train set using the online feature extractor module. A window-based normalization technique was applied to those features. In the offline model, we scale features by normalizing using the maximum absolute value of a channel [11] before applying a sliding window approach. Since the online system has access to a limited amount of data, we normalize based on the observed window. The model uses the feature vectors with a frame size of 1 second and a window size of 7 seconds. We evaluated the model using the offline P1 postprocessor to determine the efficacy of the delayed features and the window-based normalization technique. As shown by the results of experiments 1 and 4 in Table 2, these changes give us a comparable performance to the offline model. The online event decoder module utilizes this trained model for computing probabilities for the seizure and background classes. These posteriors are then postprocessed to remove spurious detections. The online postprocessor receives and saves 8 seconds of class posteriors in a buffer for further processing. It applies multiple heuristic filters (e.g., probability threshold) to make an overall decision by combining events across the channels. These filters evaluate the average confidence, the duration of a seizure, and the channels where the seizures were observed. The postprocessor delivers the label and confidence to the visualizer. The visualizer starts to display the signal as soon as it gets access to the signal file, as shown in Figure 1 using the “Signal File” and “Visualizer” blocks. Once the visualizer receives the label and confidence for the latest epoch from the postprocessor, it overlays the decision and color codes that epoch. The visualizer uses red for seizure with the label SEIZ and green for the background class with the label BCKG. Once the streaming finishes, the system saves three files: a signal file in which the sample frames are saved in the order they were streamed, a time segmented event (TSE) file with the overall decisions and confidences, and a hypotheses (HYP) file that saves the label and confidence for each epoch. The user can plot the signal and decisions using the signal and HYP files with only the visualizer by enabling appropriate options. For comparing the performance of different stages of development, we used the test set of TUSZ v1.2.1 database. It contains 1015 EEG records of varying duration. The any-overlap performance [12] of the overall system shown in Figure 2 is 40.29% sensitivity with 5.77 FAs per 24 hours. For comparison, the previous state-of-the-art model developed on this database performed at 30.71% sensitivity with 6.77 FAs per 24 hours [3]. The individual performances of the deep learning phases are as follows: Phase 1’s (P1) performance is 39.46% sensitivity and 11.62 FAs per 24 hours, and Phase 2 detects seizures with 41.16% sensitivity and 11.69 FAs per 24 hours. We trained an LSTM model with the delayed features and the window-based normalization technique for developing the online system. Using the offline decoder and postprocessor, the model performed at 36.23% sensitivity with 9.52 FAs per 24 hours. The trained model was then evaluated with the online modules. The current performance of the overall online system is 45.80% sensitivity with 28.14 FAs per 24 hours. Table 2 summarizes the performances of these systems. The performance of the online system deviates from the offline P1 model because the online postprocessor fails to combine the events as the seizure probability fluctuates during an event. The modules in the online system add a total of 11.1 seconds of delay for processing each second of the data, as shown in Figure 3. In practice, we also count the time for loading the model and starting the visualizer block. When we consider these facts, the system consumes 15 seconds to display the first hypothesis. The system detects seizure onsets with an average latency of 15 seconds. Implementing an automatic seizure detection model in real time is not trivial. We used a variety of techniques such as the file locking mechanism, multithreading, circular buffers, real-time event decoding, and signal-decision plotting to realize the system. A video demonstrating the system is available at: https://www.isip.piconepress.com/projects/nsf_pfi_tt/resources/videos/realtime_eeg_analysis/v2.5.1/video_2.5.1.mp4. The final conference submission will include a more detailed analysis of the online performance of each module. ACKNOWLEDGMENTS Research reported in this publication was most recently supported by the National Science Foundation Partnership for Innovation award number IIP-1827565 and the Pennsylvania Commonwealth Universal Research Enhancement Program (PA CURE). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the official views of any of these organizations. REFERENCES [1] A. Craik, Y. He, and J. L. Contreras-Vidal, “Deep learning for electroencephalogram (EEG) classification tasks: a review,” J. Neural Eng., vol. 16, no. 3, p. 031001, 2019. https://doi.org/10.1088/1741-2552/ab0ab5. [2] A. C. Bridi, T. Q. Louro, and R. C. L. Da Silva, “Clinical Alarms in intensive care: implications of alarm fatigue for the safety of patients,” Rev. Lat. Am. Enfermagem, vol. 22, no. 6, p. 1034, 2014. https://doi.org/10.1590/0104-1169.3488.2513. [3] M. Golmohammadi, V. Shah, I. Obeid, and J. Picone, “Deep Learning Approaches for Automatic Seizure Detection from Scalp Electroencephalograms,” in Signal Processing in Medicine and Biology: Emerging Trends in Research and Applications, 1st ed., I. Obeid, I. Selesnick, and J. Picone, Eds. New York, New York, USA: Springer, 2020, pp. 233–274. https://doi.org/10.1007/978-3-030-36844-9_8. [4] “CFM Olympic Brainz Monitor.” [Online]. Available: https://newborncare.natus.com/products-services/newborn-care-products/newborn-brain-injury/cfm-olympic-brainz-monitor. [Accessed: 17-Jul-2020]. [5] M. L. Scheuer, S. B. Wilson, A. Antony, G. Ghearing, A. Urban, and A. I. Bagic, “Seizure Detection: Interreader Agreement and Detection Algorithm Assessments Using a Large Dataset,” J. Clin. Neurophysiol., 2020. https://doi.org/10.1097/WNP.0000000000000709. [6] A. Harati, M. Golmohammadi, S. Lopez, I. Obeid, and J. Picone, “Improved EEG Event Classification Using Differential Energy,” in Proceedings of the IEEE Signal Processing in Medicine and Biology Symposium, 2015, pp. 1–4. https://doi.org/10.1109/SPMB.2015.7405421. [7] V. Shah, C. Campbell, I. Obeid, and J. Picone, “Improved Spatio-Temporal Modeling in Automated Seizure Detection using Channel-Dependent Posteriors,” Neurocomputing, 2021. [8] W. Tatum, A. Husain, S. Benbadis, and P. Kaplan, Handbook of EEG Interpretation. New York City, New York, USA: Demos Medical Publishing, 2007. [9] D. P. Bovet and C. Marco, Understanding the Linux Kernel, 3rd ed. O’Reilly Media, Inc., 2005. https://www.oreilly.com/library/view/understanding-the-linux/0596005652/. [10] V. Shah et al., “The Temple University Hospital Seizure Detection Corpus,” Front. Neuroinform., vol. 12, pp. 1–6, 2018. https://doi.org/10.3389/fninf.2018.00083. [11] F. Pedregosa et al., “Scikit-learn: Machine Learning in Python,” J. Mach. Learn. Res., vol. 12, pp. 2825–2830, 2011. https://dl.acm.org/doi/10.5555/1953048.2078195. [12] J. Gotman, D. Flanagan, J. Zhang, and B. Rosenblatt, “Automatic seizure detection in the newborn: Methods and initial evaluation,” Electroencephalogr. Clin. Neurophysiol., vol. 103, no. 3, pp. 356–362, 1997. https://doi.org/10.1016/S0013-4694(97)00003-9. 
    more » « less
  4. The concept of long-wavelength ductile flow of lower crustal material, or channel flow, has emerged to explain the evolution of large hot orogens. In this model, growth of heat producing crust during collision leads to melt-weakening and flow of lower crust in response to tectonic forcing or long-wavelength gradients in gravitational potential energy. In the Himalayan-Tibetan (HT) orogen where the model was originally proposed, it has been hypothesized that a Miocene orogen-normal channel was active and that there was a more recent switch to orogen-parallel “escape” flow as the front of the orogen began to deform as a thrust wedge. However, because this hypothesized HT orogenic channel is largely subsurface it cannot be directly examined, making it difficult to test these hypotheses. The Inner Piedmont (IP), southern Appalachians has been proposed to be an exhumed orogenic channel based on inverted metamorphic isograds, extensive migmatization, and a large-scale curved mineral lineation pattern that is consistent with a shift from orogen-normal to orogen-parallel flow. To test the viability of the channel flow model in the IP, we construct pressure-temperature-time (P-T-t) paths and compare these to existing models which indicate that peak temperatures and residence times will differ between thrust wedge and channel flow models. The P-T-t paths are constructed using isochemical phase diagram sections (pseudosections), garnet compositions, monazite geochronology, and 40Ar/39Ar thermochronology to define prograde to retrograde conditions and residence times. The channel flow models require temperatures above 700-750°C to initiate and maintain flow. Preliminary pseudosections from the northern IP Brindle Creek fault zone indicate prograde to peak conditions of 815–820 °C and 7.9–9.3 kbar, and retrograde conditions of 720–730 °C and 5.3–5.4 kbar based on observed garnet compositions and sample mineralogy (Qtz + Pl + Bt + Sil + Grt ± Ms ± Ep ± Ilm ± Rt). Pseudosections are still being revised, however if confirmed, the P-T conditions are compatible with channel flow in the IP. Future model revisions and age data from samples forming a transect across the IP and into the adjacent Carolina superterrane and eastern Blue Ridge will be used to compare the P-T-t histories between the prdoposed channel and surrounding units. 
    more » « less
  5. Key points

    Association of plasma membrane BKCachannels with BK‐β subunits shapes their biophysical properties and physiological roles; however, functional modulation of the mitochondrial BKCachannel (mitoBKCa) by BK‐β subunits is not established.

    MitoBKCa‐α and the regulatory BK‐β1 subunit associate in mouse cardiac mitochondria.

    A large fraction of mitoBKCadisplay properties similar to that of plasma membrane BKCawhen associated with BK‐β1 (left‐shifted voltage dependence of activation,V1/2 = −55 mV, 12 µmmatrix Ca2+).

    In BK‐β1 knockout mice, cardiac mitoBKCadisplayed a lowPoand a depolarizedV1/2of activation (+47 mV at 12 µmmatrix Ca2+)

    Co‐expression of BKCawith the BK‐β1 subunit in HeLa cells doubled the density of BKCain mitochondria.

    The present study supports the view that the cardiac mitoBKCachannel is functionally modulated by the BK‐β1 subunit; proper targeting and activation of mitoBKCashapes mitochondrial Ca2+handling.

    Abstract

    Association of the plasma membrane BKCachannel with auxiliary BK‐β1–4 subunits profoundly affects the regulatory mechanisms and physiological processes in which this channel participates. However, functional association of mitochondrial BK (mitoBKCa) with regulatory subunits is unknown. We report that mitoBKCafunctionally associates with its regulatory subunit BK‐β1 in adult rodent cardiomyocytes. Cardiac mitoBKCais a calcium‐ and voltage‐activated channel that is sensitive to paxilline with a large conductance for K+of 300 pS. Additionally, mitoBKCadisplays a high open probability (Po) and voltage half‐activation (V1/2 = −55 mV,n = 7) resembling that of plasma membrane BKCawhen associated with its regulatory BK‐β1 subunit. Immunochemistry assays demonstrated an interaction between mitochondrial BKCa‐α and its BK‐β1 subunit. Mitochondria from the BK‐β1 knockout (KO) mice showed sparse mitoBKCacurrents (five patches with mitoBKCaactivity out of 28 total patches fromn = 5 different hearts), displaying a depolarizedV1/2of activation (+47 mV in 12 µmmatrix Ca2+). The reduced activity of mitoBKCawas accompanied by a high expression of BKCatranscript in the BK‐β1 KO, suggesting a lower abundance of mitoBKCachannels in this genotype. Accordingly, BK‐β1subunit increased the localization of BKDEC (i.e. the splice variant of BKCathat specifically targets mitochondria) into mitochondria by two‐fold. Importantly, both paxilline‐treated and BK‐β1 KO mitochondria displayed a more rapid Ca2+overload, featuring an early opening of the mitochondrial transition pore. We provide strong evidence that mitoBKCaassociates with its regulatory BK‐β1 subunit in cardiac mitochondria, ensuring proper targeting and activation of the mitoBKCachannel that helps to maintain mitochondrial Ca2+homeostasis.

     
    more » « less