Showing papers by "Samsung published in 2020"

Efficient and Robust Approximate Nearest Neighbor Search Using Hierarchical Navigable Small World Graphs

Abstract: We present a new approach for the approximate K-nearest neighbor search based on navigable small world graphs with controllable hierarchy (Hierarchical NSW, HNSW). The proposed solution is fully graph-based, without any need for additional search structures (typically used at the coarse search stage of the most proximity graph techniques). Hierarchical NSW incrementally builds a multi-layer structure consisting of a hierarchical set of proximity graphs (layers) for nested subsets of the stored elements. The maximum layer in which an element is present is selected randomly with an exponentially decaying probability distribution. This allows producing graphs similar to the previously studied Navigable Small World (NSW) structures while additionally having the links separated by their characteristic distance scales. Starting the search from the upper layer together with utilizing the scale separation boosts the performance compared to NSW and allows a logarithmic complexity scaling. Additional employment of a heuristic for selecting proximity graph neighbors significantly increases performance at high recall and in case of highly clustered data. Performance evaluation has demonstrated that the proposed general metric space search index is able to strongly outperform previous opensource state-of-the-art vector-only approaches. Similarity of the algorithm to the skip list structure allows straightforward balanced distributed implementation.

High-energy long-cycling all-solid-state lithium metal batteries enabled by silver–carbon composite anodes

Abstract: An all-solid-state battery with a lithium metal anode is a strong candidate for surpassing conventional lithium-ion battery capabilities. However, undesirable Li dendrite growth and low Coulombic efficiency impede their practical application. Here we report that a high-performance all-solid-state lithium metal battery with a sulfide electrolyte is enabled by a Ag–C composite anode with no excess Li. We show that the thin Ag–C layer can effectively regulate Li deposition, which leads to a genuinely long electrochemical cyclability. In our full-cell demonstrations, we employed a high-Ni layered oxide cathode with a high specific capacity (>210 mAh g−1) and high areal capacity (>6.8 mAh cm−2) and an argyrodite-type sulfide electrolyte. A warm isostatic pressing technique was also introduced to improve the contact between the electrode and the electrolyte. A prototype pouch cell (0.6 Ah) thus prepared exhibited a high energy density (>900 Wh l−1), stable Coulombic efficiency over 99.8% and long cycle life (1,000 times). Solid-state Li metal batteries represent one of the most promising rechargeable battery technologies. Here the authors report an exceptional high-performance prototype solid-state pouch cell made of a sulfide electrolyte, a high-Ni layered oxide cathode and, in particular, a silver–carbon composite anode with no excess Li.

Understanding interface stability in solid-state batteries

Abstract: Solid-state batteries (SSBs) using a solid electrolyte show potential for providing improved safety as well as higher energy and power density compared with conventional Li-ion batteries. However, two critical bottlenecks remain: the development of solid electrolytes with ionic conductivities comparable to or higher than those of conventional liquid electrolytes and the creation of stable interfaces between SSB components, including the active material, solid electrolyte and conductive additives. Although the first goal has been achieved in several solid ionic conductors, the high impedance at various solid/solid interfaces remains a challenge. Recently, computational models based on ab initio calculations have successfully predicted the stability of solid electrolytes in various systems. In addition, a large amount of experimental data has been accumulated for different interfaces in SSBs. In this Review, we summarize the experimental findings for various classes of solid electrolytes and relate them to computational predictions, with the aim of providing a deeper understanding of the interfacial reactions and insight for the future design and engineering of interfaces in SSBs. We find that, in general, the electrochemical stability and interfacial reaction products can be captured with a small set of chemical and physical principles. The reliable operation of solid-state batteries requires stable or passivating interfaces between solid components. In this Review, we discuss models for interfacial reactions and relate the predictions to experimental findings, aiming to provide a deeper understanding of interface stability.

Multifunctional materials for implantable and wearable photonic healthcare devices.

Abstract: Numerous light-based diagnostic and therapeutic devices are routinely used in the clinic. These devices have a familiar look as items plugged in the wall or placed at patients' bedsides, but recently, many new ideas have been proposed for the realization of implantable or wearable functional devices. Many advances are being fuelled by the development of multifunctional materials for photonic healthcare devices. However, the finite depth of light penetration in the body is still a serious constraint for their clinical applications. In this Review, we discuss the basic concepts and some examples of state-of-the-art implantable and wearable photonic healthcare devices for diagnostic and therapeutic applications. First, we describe emerging multifunctional materials critical to the advent of next-generation implantable and wearable photonic healthcare devices and discuss the path for their clinical translation. Then, we examine implantable photonic healthcare devices in terms of their properties and diagnostic and therapeutic functions. We next describe exemplary cases of noninvasive, wearable photonic healthcare devices across different anatomical applications. Finally, we discuss the future research directions for the field, in particular regarding mobile healthcare and personalized medicine.

Efficient and stable blue quantum dot light-emitting diode.

Abstract: The visualization of accurate colour information using quantum dots has been explored for decades, and commercial products employing environmentally friendly materials are currently available as backlights1. However, next-generation electroluminescent displays based on quantum dots require the development of an efficient and stable cadmium-free blue-light-emitting device, which has remained a challenge because of the inferior photophysical properties of blue-light-emitting materials2,3. Here we present the synthesis of ZnSe-based blue-light-emitting quantum dots with a quantum yield of unity. We found that hydrofluoric acid and zinc chloride additives are effective at enhancing luminescence efficiency by eliminating stacking faults in the ZnSe crystalline structure. In addition, chloride passivation through liquid or solid ligand exchange leads to slow radiative recombination, high thermal stability and efficient charge-transport properties. We constructed double quantum dot emitting layers with gradient chloride content in a light-emitting diode to facilitate hole transport, and the resulting device showed an efficiency at the theoretical limit, high brightness and long operational lifetime. We anticipate that our efficient and stable blue quantum dot light-emitting devices can facilitate the development of electroluminescent full-colour displays using quantum dots. Cadmium-free blue quantum dot light-emitting diodes are constructed with a quantum yield of unity, an efficiency at the theoretical limit, high brightness and long operational lifetime.

Voltage decay and redox asymmetry mitigation by reversible cation migration in lithium-rich layered oxide electrodes

Abstract: Despite the high energy density of lithium-rich layered-oxide electrodes, their real-world implementation in batteries is hindered by the substantial voltage decay on cycling. This voltage decay is widely accepted to mainly originate from progressive structural rearrangements involving irreversible transition-metal migration. As prevention of this spontaneous cation migration has proven difficult, a paradigm shift toward management of its reversibility is needed. Herein, we demonstrate that the reversibility of the cation migration of lithium-rich nickel manganese oxides can be remarkably improved by altering the oxygen stacking sequences in the layered structure and thereby dramatically reducing the voltage decay. The preeminent intra-cycle reversibility of the cation migration is experimentally visualized, and first-principles calculations reveal that an O2-type structure restricts the movements of transition metals within the Li layer, which effectively streamlines the returning migration path of the transition metals. Furthermore, we propose that the enhanced reversibility mitigates the asymmetry of the anionic redox in conventional lithium-rich electrodes, promoting the high-potential anionic reduction, thereby reducing the subsequent voltage hysteresis. Our findings demonstrate that regulating the reversibility of the cation migration is a practical strategy to reduce voltage decay and hysteresis in lithium-rich layered materials.

Generalized ODIN: Detecting Out-of-distribution Image without Learning from Out-of-distribution Data

Abstract: Deep neural networks have attained remarkable performance when applied to data that comes from the same distribution as that of the training set, but can significantly degrade otherwise. Therefore, detecting whether an example is out-of-distribution (OoD) is crucial to enable a system that can reject such samples or alert users. Recent works have made significant progress on OoD benchmarks consisting of small image datasets. However, many recent methods based on neural networks rely on training or tuning with both in-distribution and out-of-distribution data. The latter is generally hard to define a-priori, and its selection can easily bias the learning. We base our work on a popular method ODIN, proposing two strategies for freeing it from the needs of tuning with OoD data, while improving its OoD detection performance. We specifically propose to decompose confidence scoring as well as a modified input pre-processing method. We show that both of these significantly help in detection performance. Our further analysis on a larger scale image dataset shows that the two types of distribution shifts, specifically semantic shift and non-semantic shift, present a significant difference in the difficulty of the problem, providing an analysis of when ODIN-like strategies do or do not work.

Capivasertib Plus Paclitaxel Versus Placebo Plus Paclitaxel As First-Line Therapy for Metastatic Triple-Negative Breast Cancer: The PAKT Trial.

Abstract: PURPOSEThe phosphatidylinositol 3-kinase (PI3K)/AKT signaling pathway is frequently activated in triple-negative breast cancer (TNBC). The AKT inhibitor capivasertib has shown preclinical activity ...

Artificial Intelligence-Enabled Cellular Networks: A Critical Path to Beyond-5G and 6G

Abstract: Mobile network operators (MNOs) are in the process of overlaying their conventional macro cellular networks with shorter range cells such as outdoor pico cells. The resultant increase in network complexity creates substantial overhead in terms of operating expenses, time, and labor for their planning and management. Artificial intelligence (AI) offers the potential for MNOs to operate their networks in a more organic and cost-efficient manner. We argue that deploying AI in fifth generation (5G) and beyond will require surmounting significant technical barriers in terms of robustness, performance, and complexity. We outline future research directions, identify top five challenges, and present a possible roadmap to realize the vision of AI-enabled cellular networks for Beyond- 5G and sixth generation (6G) networks.

Verification, analytical validation, and clinical validation (V3): the foundation of determining fit-for-purpose for Biometric Monitoring Technologies (BioMeTs).

Abstract: Digital medicine is an interdisciplinary field, drawing together stakeholders with expertize in engineering, manufacturing, clinical science, data science, biostatistics, regulatory science, ethics, patient advocacy, and healthcare policy, to name a few. Although this diversity is undoubtedly valuable, it can lead to confusion regarding terminology and best practices. There are many instances, as we detail in this paper, where a single term is used by different groups to mean different things, as well as cases where multiple terms are used to describe essentially the same concept. Our intent is to clarify core terminology and best practices for the evaluation of Biometric Monitoring Technologies (BioMeTs), without unnecessarily introducing new terms. We focus on the evaluation of BioMeTs as fit-for-purpose for use in clinical trials. However, our intent is for this framework to be instructional to all users of digital measurement tools, regardless of setting or intended use. We propose and describe a three-component framework intended to provide a foundational evaluation framework for BioMeTs. This framework includes (1) verification, (2) analytical validation, and (3) clinical validation. We aim for this common vocabulary to enable more effective communication and collaboration, generate a common and meaningful evidence base for BioMeTs, and improve the accessibility of the digital medicine field.

Generalized ODIN: Detecting Out-of-Distribution Image Without Learning From Out-of-Distribution Data

Abstract: Deep neural networks have attained remarkable performance when applied to data that comes from the same distribution as that of the training set, but can significantly degrade otherwise. Therefore, detecting whether an example is out-of-distribution (OoD) is crucial to enable a system that can reject such samples or alert users. Recent works have made significant progress on OoD benchmarks consisting of small image datasets. However, many recent methods based on neural networks rely on training or tuning with both in-distribution and out-of-distribution data. The latter is generally hard to define a-priori, and its selection can easily bias the learning. We base our work on a popular method ODIN, proposing two strategies for freeing it from the needs of tuning with OoD data, while improving its OoD detection performance. We specifically propose to decompose confidence scoring as well as a modified input pre-processing method. We show that both of these significantly help in detection performance. Our further analysis on a larger scale image dataset shows that the two types of distribution shifts, specifically semantic shift and non-semantic shift, present a significant difference in the difficulty of the problem, providing an analysis of when ODIN-like strategies do or do not work.

Incremental Few-Shot Object Detection

Abstract: Existing object detection methods typically rely on the availability of abundant labelled training samples per class and offline model training in a batch mode. These requirements substantially limit their scalability to open-ended accommodation of novel classes with limited labelled training data, both in terms of model accuracy and training efficiency during deployment. We present the first study aiming to go beyond these limitations by considering the Incremental Few-Shot Detection (iFSD) problem setting, where new classes must be registered incrementally (without revisiting base classes) and with few examples. To this end we propose OpeN-ended Centre nEt (ONCE), a detector designed for incrementally learning to detect novel class objects with few examples. This is achieved by an elegant adaptation of the efficient CentreNet detector to the few-shot learning scenario, and meta-learning a class-wise code generator model for registering novel classes. ONCE fully respects the incremental learning paradigm, with novel class registration requiring only a single forward pass of few-shot training samples, and no access to base classes – thus making it suitable for deployment on embedded devices, etc. Extensive experiments conducted on both the standard object detection (COCO, PASCAL VOC) and fashion landmark detection (DeepFashion2) tasks show the feasibility of iFSD for the first time, opening an interesting and very important line of research.

Metasurface-driven OLED displays beyond 10,000 pixels per inch

Abstract: Optical metasurfaces are starting to find their way into integrated devices, where they can enhance and control the emission, modulation, dynamic shaping, and detection of light waves. In this study, we show that the architecture of organic light-emitting diode (OLED) displays can be completely reenvisioned through the introduction of nanopatterned metasurface mirrors. In the resulting meta-OLED displays, different metasurface patterns define red, green, and blue pixels and ensure optimized extraction of these colors from organic, white light emitters. This new architecture facilitates the creation of devices at the ultrahigh pixel densities (>10,000 pixels per inch) required in emerging display applications (for instance, augmented reality) that use scalable nanoimprint lithography. The fabricated pixels also offer twice the luminescence efficiency and superior color purity relative to standard color-filtered white OLEDs.

Realistic Speech-Driven Facial Animation with GANs

Abstract: Speech-driven facial animation is the process that automatically synthesizes talking characters based on speech signals. The majority of work in this domain creates a mapping from audio features to visual features. This approach often requires post-processing using computer graphics techniques to produce realistic albeit subject dependent results. We present an end-to-end system that generates videos of a talking head, using only a still image of a person and an audio clip containing speech, without relying on handcrafted intermediate features. Our method generates videos which have (a) lip movements that are in sync with the audio and (b) natural facial expressions such as blinks and eyebrow movements. Our temporal GAN uses 3 discriminators focused on achieving detailed frames, audio-visual synchronization, and realistic expressions. We quantify the contribution of each component in our model using an ablation study and we provide insights into the latent representation of the model. The generated videos are evaluated based on sharpness, reconstruction quality, lip-reading accuracy, synchronization as well as their ability to generate natural blinks.

The Eighth Visual Object Tracking VOT2020 Challenge Results

Abstract: The Visual Object Tracking challenge VOT2020 is the eighth annual tracker benchmarking activity organized by the VOT initiative. Results of 58 trackers are presented; many are state-of-the-art trackers published at major computer vision conferences or in journals in the recent years. The VOT2020 challenge was composed of five sub-challenges focusing on different tracking domains: (i) VOT-ST2020 challenge focused on short-term tracking in RGB, (ii) VOT-RT2020 challenge focused on “real-time” short-term tracking in RGB, (iii) VOT-LT2020 focused on long-term tracking namely coping with target disappearance and reappearance, (iv) VOT-RGBT2020 challenge focused on short-term tracking in RGB and thermal imagery and (v) VOT-RGBD2020 challenge focused on long-term tracking in RGB and depth imagery. Only the VOT-ST2020 datasets were refreshed. A significant novelty is introduction of a new VOT short-term tracking evaluation methodology, and introduction of segmentation ground truth in the VOT-ST2020 challenge – bounding boxes will no longer be used in the VOT-ST challenges. A new VOT Python toolkit that implements all these novelites was introduced. Performance of the tested trackers typically by far exceeds standard baselines. The source code for most of the trackers is publicly available from the VOT page. The dataset, the evaluation kit and the results are publicly available at the challenge website (http://votchallenge.net).

High Active Material Loading in All‐Solid‐State Battery Electrode via Particle Size Optimization

Abstract: Author(s): Shi, T; Tu, Q; Tian, Y; Xiao, Y; Miara, LJ; Kononova, O; Ceder, G | Abstract: Low active material loading in the composite electrode of all-solid-state batteries (SSBs) is one of the main reasons for the low energy density in current SSBs. In this work, it is demonstrated with both modeling and experiments that in the regime of high cathode loading, the utilization of cathode material in the solid-state composite is highly dependent on the particle size ratio of the cathode to the solid-state conductor. The modeling, confirmed by experimental data, shows that higher cathode loading and therefore an increased energy density can be achieved by increasing the ratio of the cathode to conductor particle size. These results are consistent with ionic percolation being the limiting factor in cold-pressed solid-state cathode materials and provide specific guidelines on how to improve the energy density of composite cathodes for solid-state batteries. By reducing solid electrolyte particle size and increasing the cathode active material particle size, over 50 vol% cathode active material loading with high cathode utilization is able to be experimentally achieved, demonstrating that a commercially-relevant, energy-dense cathode composite is achievable through simple mixing and pressing method.

Fast sulfate formation from oxidation of SO 2 by NO 2 and HONO observed in Beijing haze

Abstract: Severe events of wintertime particulate air pollution in Beijing (winter haze) are associated with high relative humidity (RH) and fast production of particulate sulfate from the oxidation of sulfur dioxide (SO2) emitted by coal combustion. There has been considerable debate regarding the mechanism for SO2 oxidation. Here we show evidence from field observations of a haze event that rapid oxidation of SO2 by nitrogen dioxide (NO2) and nitrous acid (HONO) takes place, the latter producing nitrous oxide (N2O). Sulfate shifts to larger particle sizes during the event, indicative of fog/cloud processing. Fog and cloud readily form under winter haze conditions, leading to high liquid water contents with high pH (>5.5) from elevated ammonia. Such conditions enable fast aqueous-phase oxidation of SO2 by NO2, producing HONO which can in turn oxidize SO2 to yield N2O.This mechanism could provide an explanation for sulfate formation under some winter haze conditions.

Ultralow-dielectric-constant amorphous boron nitride

Abstract: Decrease in processing speed due to increased resistance and capacitance delay is a major obstacle for the down-scaling of electronics1-3 Minimizing the dimensions of interconnects (metal wires that connect different electronic components on a chip) is crucial for the miniaturization of devices Interconnects are isolated from each other by non-conducting (dielectric) layers So far, research has mostly focused on decreasing the resistance of scaled interconnects because integration of dielectrics using low-temperature deposition processes compatible with complementary metal-oxide-semiconductors is technically challenging Interconnect isolation materials must have low relative dielectric constants (κ values), serve as diffusion barriers against the migration of metal into semiconductors, and be thermally, chemically and mechanically stable Specifically, the International Roadmap for Devices and Systems recommends4 the development of dielectrics with κ values of less than 2 by 2028 Existing low-κ materials (such as silicon oxide derivatives, organic compounds and aerogels) have κ values greater than 2 and poor thermo-mechanical properties5 Here we report three-nanometre-thick amorphous boron nitride films with ultralow κ values of 178 and 116 (close to that of air, κ = 1) at operation frequencies of 100 kilohertz and 1 megahertz, respectively The films are mechanically and electrically robust, with a breakdown strength of 73 megavolts per centimetre, which exceeds requirements Cross-sectional imaging reveals that amorphous boron nitride prevents the diffusion of cobalt atoms into silicon under very harsh conditions, in contrast to reference barriers Our results demonstrate that amorphous boron nitride has excellent low-κ dielectric characteristics for high-performance electronics

Class-incremental Learning via Deep Model Consolidation

Abstract: Deep neural networks (DNNs) often suffer from "catastrophic forgetting" during incremental learning (IL) — an abrupt degradation of performance on the original set of classes when the training objective is adapted to a newly added set of classes. Existing IL approaches tend to produce a model that is biased towards either the old classes or new classes, unless with the help of exemplars of the old data. To address this issue, we propose a class-incremental learning paradigm called Deep Model Consolidation (DMC), which works well even when the original training data is not available. The idea is to first train a separate model only for the new classes, and then combine the two individual models trained on data of two distinct set of classes (old classes and new classes) via a novel double distillation training objective. The two existing models are consolidated by exploiting publicly available unlabeled auxiliary data. This overcomes the potential difficulties due to unavailability of original training data. Compared to the state-of-the-art techniques, DMC demonstrates significantly better performance in image classification (CIFAR-100 and CUB-200) and object detection (PASCAL VOC 2007) in the single-headed IL setting.

U-GAT-IT: Unsupervised Generative Attentional Networks with Adaptive Layer-Instance Normalization for Image-to-Image Translation

Abstract: We propose a novel method for unsupervised image-to-image translation, which incorporates a new attention module and a new learnable normalization function in an end-to-end manner. The attention module guides our model to focus on more important regions distinguishing between source and target domains based on the attention map obtained by the auxiliary classifier. Unlike previous attention-based method which cannot handle the geometric changes between domains, our model can translate both images requiring holistic changes and images requiring large shape changes. Moreover, our new AdaLIN (Adaptive Layer-Instance Normalization) function helps our attention-guided model to flexibly control the amount of change in shape and texture by learned parameters depending on datasets. Experimental results show the superiority of the proposed method compared to the existing state-of-the-art models with a fixed network architecture and hyper-parameters.

FedFast: Going Beyond Average for Faster Training of Federated Recommender Systems

Abstract: Federated learning (FL) is quickly becoming the de facto standard for the distributed training of deep recommendation models, using on-device user data and reducing server costs. In a typical FL process, a central server tasks end-users to train a shared recommendation model using their local data. The local models are trained over several rounds on the users' devices and the server combines them into a global model, which is sent to the devices for the purpose of providing recommendations. Standard FL approaches use randomly selected users for training at each round, and simply average their local models to compute the global model. The resulting federated recommendation models require significant client effort to train and many communication rounds before they converge to a satisfactory accuracy. Users are left with poor quality recommendations until the late stages of training. We present a novel technique, FedFast, to accelerate distributed learning which achieves good accuracy for all users very early in the training process. We achieve this by sampling from a diverse set of participating clients in each training round and applying an active aggregation method that propagates the updated model to the other clients. Consequently, with FedFast the users benefit from far lower communication costs and more accurate models that can be consumed anytime during the training process even at the very early stages. We demonstrate the efficacy of our approach across a variety of benchmark datasets and in comparison to state-of-the-art recommendation techniques.

F-BRS: Rethinking Backpropagating Refinement for Interactive Segmentation

Abstract: Deep neural networks have become a mainstream approach to interactive segmentation. As we show in our experiments, while for some images a trained network provides accurate segmentation result with just a few clicks, for some unknown objects it cannot achieve satisfactory result even with a large amount of user input. Recently proposed backpropagating refinement scheme (BRS) introduces an optimization problem for interactive segmentation that results in significantly better performance for the hard cases. At the same time, BRS requires running forward and backward pass through a deep network several times that leads to significantly increased computational budget per click compared to other methods. We propose f-BRS (feature backpropagating refinement scheme) that solves an optimization problem with respect to auxiliary variables instead of the network inputs, and requires running forward and backward passes just for a small part of a network. Experiments on GrabCut, Berkeley, DAVIS and SBD datasets set new state-of-the-art at an order of magnitude lower time per click compared to original BRS. The code and trained models are available at https://github.com/saic-vul/fbrs_interactive_segmentation.

Non-Stationarities in Extra-Large-Scale Massive MIMO

Abstract: Massive MIMO, a key technology for increasing area spectral efficiency in cellular systems, was developed assuming moderately sized apertures In this article, we argue that massive MIMO systems behave differently in large-scale regimes due to spatial non-stationarity In the large-scale regime, with arrays of around 50 wavelengths, the terminals see the whole array but non-stationarities occur because different regions of the array see different propagation paths At even larger dimensions, which we call the extra-large scale regime, terminals see a portion of the array and inside the first type of non-stationarities might occur We show that the non-stationarity properties of the massive MIMO channel change several important MIMO design aspects In simulations, we demonstrate how non-stationarity is a curse when neglected but a blessing when embraced in terms of computational load and multi-user transceiver design

Environmentally Friendly InP-based Quantum Dots for Efficient Wide Color Gamut Displays

Abstract: Quantum dots (QD) are regarded as ideal light emitters for current and next-generation displays. Hence, there is an urgent need to produce environmentally friendly QDs that show high efficiency and...

Training binary neural networks with real-to-binary convolutions

Abstract: This paper shows how to train binary networks to within a few percent points (~3-5 %) of the full precision counterpart with a negligible increase in the computational cost. In particular, we first show how to build a strong baseline, which already achieves state-of-the-art accuracy, by combining recently proposed advances, and carefully tuning the optimization procedure. Secondly, we show that by attempting to minimize the discrepancy between the output of the binary and the corresponding real-valued convolution additional significant accuracy gains can be obtained. We materialize this idea in two complementary ways: (1) with a loss function, during training, by matching the spatial attention maps computed at the output of the binary and real-valued convolutions, and (2) in data-driven manner, by using the real-valued activations being available during inference prior to the binarization process for re-scaling the activations right after the binary convolution. Finally, we show that, when putting all of our improvements together, the resulting model reduces the gap to its real-valued counterpart to less than 3% and 5% top-1 error on CIFAR-100 and ImageNet, respectively, when using a ResNet-18 architecture.

Deep Learning-Based mmWave Beam Selection for 5G NR/6G With Sub-6 GHz Channel Information: Algorithms and Prototype Validation

Abstract: In fifth-generation (5G) communications, millimeter wave (mmWave) is one of the key technologies to increase the data rate. To overcome this technology's poor propagation characteristics, it is necessary to employ a number of antennas and form narrow beams. It becomes crucial then, especially for initial access, to attain fine beam alignment between a next generation NodeB (gNB) and a user equipment (UE). The current 5G New Radio (NR) standard, however, adopts an exhaustive search-based beam sweeping, which causes time overhead of a half frame for initial beam establishment. In this paper, we propose a deep learning-based beam selection, which is compatible with the 5G NR standard. To select a mmWave beam, we exploit sub-6 GHz channel information. We introduce a deep neural network (DNN) structure and explain how we estimate a power delay profile (PDP) of a sub-6 GHz channel, which is used as an input of the DNN. We then validate its performance with real environment-based 3D ray-tracing simulations and over-the-air experiments with a mmWave prototype. Evaluation results confirm that, with support from the sub-6 GHz connection, the proposed beam selection reduces the beam sweeping overhead by up to 79.3 %.

Lipreading Using Temporal Convolutional Networks

Abstract: Lip-reading has attracted a lot of research attention lately thanks to advances in deep learning. The current state-of-the-art model for recognition of isolated words in-the-wild consists of a residual network and Bidirectional Gated Recurrent Unit (BGRU) layers. In this work, we address the limitations of this model and we propose changes which further improve its performance. Firstly, the BGRU layers are replaced with Temporal Convolutional Networks (TCN). Secondly, we greatly simplify the training procedure, which allows us to train the model in one single stage. Thirdly, we show that the current state-of-the-art methodology produces models that do not generalize well to variations on the sequence length, and we address this issue by proposing a variable-length augmentation. We present results on the largest publicly-available datasets for isolated word recognition in English and Mandarin, LRW and LRW1000, respectively. Our proposed model1 results in an absolute improvement of 1.2% and 3.2%, respectively, in these datasets which is the new state-of-the-art performance.