Modern face alignment methods have become quite accurate at predicting the locations of facial landmarks, but they do not typically estimate uncertainty their predicted nor predict whether landmarks are visible. In this paper, we present a novel framework for jointly landmark locations, associated uncertainties these and visibilities. We model as mixed random variables them using deep network trained our proposed Location, Uncertainty, Visibility Likelihood (LUVLi) loss. addition, release an...

Discovering and exploiting shared, invariant neural activity in electroencephalogram (EEG) based classification tasks is of significant interest for generalizability decoding models across subjects or EEG recording sessions. While deep networks are recently emerging as generic feature extractors, this transfer learning aspect usually relies on the prior assumption that naturally behave subject- (or session-) extractors. We propose a further step towards invariance frameworks systemic way...

Abstract A novel conditional variational autoencoder (CVAE) model for designing nanopatterned integrated photonic components is proposed. In particular, it shown that prediction capability of the CVAE can be significantly improved by adversarial censoring and active learning. Moreover, generation power splitters with arbitrary splitting ratios 550 nm broadband optical responses from 1250 to 1800 are demonstrated. Nanopatterned footprints 2.25 × m 2 20 etch hole positions design space, each...

Convolutional neural networks (CNN) have been frequently used to extract subject-invariant features from electroencephalogram (EEG) for classification tasks. This approach holds the underlying assumption that electrodes are equidistant analogous pixels of an image and hence fails explore/exploit complex functional connectivity between different electrode sites. We overcome this limitation by tailoring concepts convolution pooling applied 2D grid-like inputs network Furthermore, we develop...

BIOMETRICS are an important and widely used class of methods for identity verification access control. Biometrics attractive because they inherent properties individual. They need not be remembered like passwords, easily lost or forged identifying documents. At the same time, bio- metrics fundamentally noisy irreplaceable. There always slight variations among measurements a given biometric, and, unlike passwords identification numbers, biometrics derived from physical characteristics that...

With increasing penetrations of wind generation, based on power‐electronic converters, power systems are transitioning away from well‐understood synchronous generator‐based systems, with growing implications for their stability. Issues concern will vary system size, penetration level, geographical distribution and turbine type, network topology, electricity market structure, unit commitment procedures, other factors. However, variable‐speed turbines, both onshore connected offshore through...

In this paper, we propose a hand pose estimation approach from low cost surface electromyogram (sEMG) signals using recurrent neural networks (RNN). We use the Leap Motion sensor to capture joint kinematics and Myo collect sEMG while user is performing simple finger movements. aim at building an accurate regression model that predicts features. RNN with long short-term memory (LSTM) cells account for non-linear relationship between two domains (sEMG pose). Additionally, add Gaussian mixture...

We introduce adversarial neural networks for representation learning as a novel approach to transfer in brain-computer interfaces (BCIs). The proposed aims learn subject-invariant representations by simultaneously training conditional variational autoencoder (cVAE) and an network. use shallow convolutional architectures realize the cVAE, learned encoder is transferred extract features from unseen BCI users' data decoding. demonstrate proof-of-concept of our based on analyses...

Deep learning methods for person identification based on electroencephalographic (EEG) brain activity encounters the problem of exploiting temporally correlated structures or recording session specific variability within EEG. Furthermore, recent have mostly trained and evaluated single EEG data. We address this from an invariant representation perspective. propose adversarial inference approach to extend such deep models learn session-invariant person-discriminative representations that can...

The increasing deployment of battery storage applications in both grid and electric vehicle fields is generating a vast used market. These batteries are typically recycled but could be reused second-life applications. One the challenges to obtain an accurate remaining useful life (RUL) estimation algorithm, which determines whether suitable for reuse estimates number cycles will last. In this article, RUL problem considered. We propose several health indicators (HIs), some have not been...

We analyze brain waves acquired through a consumer-grade EEG device to investigate its capabilities for user identification and authentication. First, we show the statistical significance of P300 component in event-related potential (ERP) data from 14-channel EEGs across 25 subjects. then apply variety machine learning techniques, comparing performance various different combinations dimensionality reduction technique followed by classification algorithm. Experimental results that an accuracy...

Recently, data-driven approaches motivated by modern deep learning have been applied to optical communications in place of traditional model-based counterparts.The application neural networks (DNN) allows flexible statistical analysis complicated fiber-optic systems without relying on any specific physical models.Due the inherent nonlinearity DNN, various equalizers based DNN shown significant potentials mitigate fiber nonlinearity.In this paper, we propose a turbo equalization (TEQ) as new...

Deep learning is now playing a major role in designing photonic devices, including nanostructured photonics. In this article, we investigate three models for nanophonic power splitters with multiple splitting ratios. The first model forward regression model, wherein the trained deep neural network (DNN) used within optimization loop. second an inverse which DNN constructs structure desired target performance given as input. third generative network, can randomly produce series of optimized...

We study brain-computer interfaces (BCI) based on the decoding of motor imagery (MI) from electroencephalography (EEG) neuromonitoring. The robustness MI-BCI is a major concern in practical applications, and hence various efforts literature have been made to enhance MI classification accuracy EEG signals. Recently, classifiers convolutional neural networks (CNN) achieved state-of-the-art performance. In further exploration applying CNNs data, we propose spatial component-wise network...

Motivated by the recent advancement of quantum processors, we investigate approximate optimization algorithm (QAOA) to employ quasi-maximum-likelihood (ML) decoding classical channel codes. QAOA is a hybrid quantum-classical variational algorithm, which advantageous for near-term noisy intermediate-scale (NISQ) devices, where fidelity gates limited noise and de-coherence. We first describe how construct Ising Hamiltonian model realize quasi-ML with QAOA. For level-1 QAOA, derive systematic...

We review advancements in silicon photonic (SiPh) devices and integrated circuits (SiPICs) to enable high density, low power, multi-Tb/s optical solutions for next-generation Ethernet networking compute connectivity.

This paper studies a new application of deep learning (DL) for optimizing constellations in two-way relaying with physical-layer network coding (PNC), where neural (DNN)-based modulation and demodulation are employed at each terminal relay node. We train DNNs such that the cross entropy loss is directly minimized, thus it maximizes likelihood, rather than considering Euclidean distance constellations. The proposed scheme can be extended to higher level slight modification DNN structure....

In this work we present the design and performance of a high bandwidth, low power, latency, high-density Silicon Photonic integrated circuit (SiPIC) for compute interconnects. The SiPIC has more than 4 Tbps bandwidth over eight standard single mode fiber pairs. Each pair carries DWDM transmit receive channels operating at 32 Gbps NRZ. We integrate all photonic components functions including lasers, semiconductor optical amplifiers (SOAs), modulators, photodetectors, spot size convertors...

Next-generation fiber-optic communications call for ultra-reliable forward error correction codes that are capable of low-power and low-latency decoding. In this paper, we propose a new class polar codes, whose polarization units irregularly pruned to reduce computational complexity decoding latency without sacrificing performance. We then experimentally demonstrate the proposed irregular can outperform state-of-the-art low-density parity-check (LDPC) while be reduced by at least 30% 70%,...

Beyond data communications, commercial-off-the-shelf Wi-Fi devices can be used to monitor human activities, track device locomotion, and sense the ambient environment. In particular, spatial beam attributes that are inherently available in 60-GHz IEEE 802.11ad/ay standards have shown effective terms of overhead channel measurement granularity for these indoor sensing tasks. this paper, we investigate transfer learning mitigate domain shift monitoring tasks when settings environments change...

Machine-to-machine (M2M) communications play an important role for applications that involve connections between a massive number of heterogeneous devices in home and industrial networks. For M2M networks, realizing low latency high reliability is great importance. In this paper, we show the potential polar-coded orthogonal frequency-division multiplexing (OFDM) to fulfill those requirements. We polar codes with list decoding plus cyclic redundancy check (CRC) can outperform state-of-the-art...

We show that polar codes with list+CRC decoding can outperform state-of-the-art LDPC in short block lengths. In addition, we introduce an efficient interleaver for polar-coded high-order modulations, achieving greater than 0.5 dB gain 256QAM.

Graph neural networks (GNN) are an emerging framework in the deep learning community. In most GNN applications, graph topology of data samples is provided dataset. Specifically, shift operator (GSO), which could be adjacency, Laplacian, or their normalizations, known a priori. However we often have no knowledge grand-truth underlying real-world datasets. One example this to extract subject-invariant features from physiological electroencephalogram (EEG) predict cognitive task. Previous...