Performance Ceiling Research Articles

ConViT: improving vision transformers with soft convolutional inductive biases* *This article is an updated version of: D’Ascoli S, Touvron H, Leavitt M L, Morcos A S, Biroli G and Sagun L 2021 ConViT: improving vision transformers with soft convolutional inductive biases Proc. 38th Int. Conf. Machine Learning vol 139 ed M Meila and T Zhang pp 2286–96.

Convolutional architectures have proven to be extremely successful for vision tasks. Their hard inductive biases enable sample-efficient learning, but come at the cost of a potentially lower performance ceiling. Vision transformers rely on more flexible self-attention layers, and have recently outperformed CNNs for image classification. However, they require costly pre-training on large external datasets or distillation from pre-trained convolutional networks. In this paper, we ask the following question: is it possible to combine the strengths of these two architectures while avoiding their respective limitations? To this end, we introduce gated positional self-attention (GPSA), a form of positional self-attention which can be equipped with a ‘soft’ convolutional inductive bias. We initialize the GPSA layers to mimic the locality of convolutional layers, then give each attention head the freedom to escape locality by adjusting a gating parameter regulating the attention paid to position versus content information. The resulting convolutional-like ViT architecture, ConViT, outperforms the DeiT (Touvron et al 2020 arXiv:2012.12877) on ImageNet, while offering a much improved sample efficiency. We further investigate the role of locality in learning by first quantifying how it is encouraged in vanilla self-attention layers, then analyzing how it has escaped in GPSA layers. We conclude by presenting various ablations to better understand the success of the ConViT. Our code and models are released publicly at https://github.com/facebookresearch/convit.

Physicochemical QSAR analysis of hERG inhibition revisited: towards a quantitative potency prediction.

In an earlier study (Didziapetris R & Lanevskij K (2016). J Comput Aided Mol Des. 30:1175-1188) we collected a database of publicly available hERG inhibition data for almost 6700 drug-like molecules and built a probabilistic Gradient Boosting classifier with a minimal set of physicochemical descriptors (log P, pKa, molecular size and topology parameters). This approach favored interpretability over statistical performance but still achieved an overall classification accuracy of 75%. In the current follow-up work we expanded the database (provided in Supplementary Information) to almost 9400 molecules and performed temporal validation of the model on a set of novel chemicals from recently published lead optimization projects. Validation results showed almost no performance degradation compared to the original study. Additionally, we rebuilt the model using AFT (Accelerated Failure Time) learning objective in XGBoost, which accepts both quantitative and censored data often reported in protein inhibition studies. The new model achieved a similar level of accuracy of discerning hERG blockers from non-blockers at 10 µM threshold, which can be conceived as close to the performance ceiling for methods aiming to describe only non-specific ligand interactions with hERG. Yet, this model outputs quantitative potency values (IC50) and is not tied to a particular classification cut-off. pIC50 from patch-clamp measurements can be predicted with R2 ≈ 0.4 and MAE < 0.5, which enables ligand ranking according to their expected potency levels. The employed approach can be valuable for quantitative modeling of various ADME and drug safety endpoints with a high prevalence of censored data.

Cold-start active learning for image classification

Active learning (AL) aims to select valuable samples for labeling from an unlabeled sample pool to build a training dataset with minimal annotation cost. Traditional methods always require partially and initially labeled samples to start active selection and then query annotations of samples incrementally through several iterations. However, this scheme is not effective in the deep learning scenario. On the one hand, initially labeled sample sets are not always available in the beginning. On the other hand, the performance of the traditional model is usually poor in the early iterations due to limited training feedback. For the first time, we propose a cold-start AL model based on representative (CALR) sampling, which selects valuable samples without the need for an initial labeled set or the iterative feedback of the target models. Experiments on three image classification datasets, CIFAR-10, CIFAR-100 and Caltech-256, showed that CALR achieved a new state-of-the-art performance of AL in cold-start settings. Especially in low annotation budget conditions, our method can achieve up to a 10% performance increase compared to traditional methods. Furthermore, CALR can be combined with warm-start methods to improve the start-up efficiency while further breaking the performance ceiling of AL, which makes CALR have a broader application scenario.

(Invited, Digital Presentation) Exploring Interfaces and Polarity to Realize Vertical III-Nitride Superjunction Devices

In recent years, there has been a surge of research and commercial interest in gallium nitride (GaN)-based devices for power conversion applications. This is largely motivated by the wide bandgap of GaN, which offers a unipolar limit of performance that is larger than that of silicon and silicon carbide. This performance ceiling, however, can be surpassed with the use of superjunction (SJ) structures, a strategy that has now been experimentally proven in both vertical Si- and SiC-based technologies. The ability to selectively dope lateral regions of a semiconductor is a requirement for SJ fabrication, with typical schemes relying on ion implantation, epitaxial regrowth or a combination of both. However, due to the challenges associated with applying conventional selective area doping techniques to GaN, vertical GaN SJ structures have remained elusive.Recently, our team proposed the use of lateral polar junctions (LPJs) to form vertical GaN SJ devices. This approach exploits the natural doping asymmetry between the N-polar and Ga-polar crystal orientations to simultaneously grow N-polar GaN for the n-type pillars and Ga-polar GaN for the p-type pillars. The first part of this talk will present a design framework for GaN SJ devices. This will reveal several critical design requirements: 1) doping in N-polar GaN must be reduced below 5 × 1017 cm-3 to achieve kV-class devices, 2) charge must be tightly balanced between neighboring pillars in the SJ structure to obtain efficient performance and 3) the barrier height to the N-polar GaN region must be tuned to suppress leakage currents and avoid bipolar operation.The next part of the talk will focus on the characterization of rectifying contacts to N-polar GaN. While the analysis of Schottky barrier diodes proved the ability to reduce doping to relevant levels, the low barrier height and enhanced chemical sensitivity of N-polar GaN make it challenging to process adequate devices. Thus, low-pressure chemical vapor deposited (LPCVD) silicon nitride (SiN) interlayers were introduced to increase the barrier height. This approach is enabled by surface-termination dependent barrier height and an amphoteric miniband for enhanced conduction via the interlayer. Moreover, the use of the LPCVD SiN interlayer enables operation of the diodes up to 400 °C compared to less than 200 °C for the reference case. In addition, N-polar GaN camel diodes were also designed, fabricated and tested to better tune the barrier height. The latter structure was incorporated into the first charge-balanced GaN superjunction device. A detailed electrical analysis of the GaN LPJ device will represent the final portion of this talk.The successful demonstration of these experimental building blocks paves the way for the GaN LPJ to be used in future high-voltage GaN SJ devices.

Insights into the enhanced effect of Low-Intensity Ultrasound on anammox granular sludge by relieving the embolism

Like the lethal cardiovascular embolism in humans, the serious blockage of the mass transfer channel can inactivate anammox granular sludge (AnGS). Also like the therapeutic thrombolysis in humans, the remediatory dredging of the mass transfer channel can sustain the function of AnGS. In this study, the ultrasound was thus explored to remediate the embolism of AnGS. It turned out that Low-Intensity Ultrasound (LIU) with an energy density of 5.6 kJ/(g-VSS·cm2) was optimal to achieve an obvious increase in specific anammox activity of AnGS by 34.8% to 420.3 mg-N/(g-VSS·d). The improved mass transfer of AnGS and cell membrane permeability of anammox bacteria (AnAOB) was revealed to mainly contribute to the enhanced activity. After treatment with LIU under optimal conditions, the mass transfer efficiency of AnGS was improved by 26.9% with enhanced permeability, and the proportions of pores larger than 50 μm in the surface layer of AnGS increased by 13.7% evidenced by the micro-CT observation. Correspondingly, the permeability of AnAOB was increased by inducing instantaneous pores on the cell membrane, thus stimulating the metabolism by 6.9%. Further analysis demonstrated that LIU was capable of breaking the hydrogen bonds of protein in the EPS and loosening its spatial structure, which would benefit the partially exfoliation and dissolution of EPS and resulted in enlarging and dredging pores, thus relieving the embolism of AnGS. Consequently, the results of this study could deepen the understanding of granules treated by ultrasound and also provide a promising technique for raising the ceiling of nitrogen removal performance by AnGS.

Accuracies of Training Labels and Machine Learning Models: Experiments on Delirium and Simulated Data.

Supervised predictive models require labeled data for training purposes. Complete and accurate labeled data is not always available, and imperfectly labeled data may need to serve as an alternative. An important question is if the accuracy of the labeled data creates a performance ceiling for the trained model. In this study, we trained several models to recognize the presence of delirium in clinical documents using data with annotations that are not completely accurate. In the external evaluation, the support vector machine model with a linear kernel performed best, achieving an area under the curve of 89.3% and accuracy of 88%, surpassing the 80% accuracy of the training sample. We then generated a set of simulated data and carried out a series of experiments which demonstrated that models trained on imperfect data can (but do not always) outperform the accuracy of the training data.

Crystalline molybdenum carbide−amorphous molybdenum oxide heterostructures: In situ surface reconfiguration and electronic states modulation for Li−S batteries

Designing heterostructure electrocatalysts synergizing highly polar component and conductive matrix hold great promise to regulate lithium polysulfides (LiPSs) toward high−performance Li−S batteries based on anchoring−diffusion−conversion mechanism. Nevertheless, performance ceiling exists, restrict severely by the significant LiPSs diffusion barrier, which inevitably decrease the conversion efficiency of LiPSs. Amorphization can not only promote the diffusion of LiPSs on catalyst surface but also induce the modulation of surface electronic states and strengthen the adsorption toward LiPSs, in comparison with crystalline counterpart. Herein, crystalline Mo2C−amorphous MoO3 heterostructure grown on nitrogen doped carbon nanosheets is meticulously designed as an efficient sulfur immobilizer and promoter through a facile oxygen plasma treatment strategy. With the integrated advantage of improved adsorption and facilitated diffusion of LiPSs by amorphous MoO3, superior LiPSs redox kinetics by the catalytic effect of conductive Mo2C, and long-range electron transfer network arising from nitrogen doped carbon nanosheets, the assembled battery based on the as-developed modified separator endows excellent electrochemical performance and satisfactory cycling stability. This work provides an easy route to synergistic regulate crystallinity and heterogeneous interfaces of multifunctional catalysts for high energy density Li−S batteries.

BoostNet: a method to enhance the performance of deep learning model on musculoskeletal radiographs X-ray images

In clinical treatment, deep learning plays a pivotal role in medical image classification. Deep learning techniques provide opportunities for radiologists and orthopedic to ease out their lives with faster and more accurate results. The traditional deep learning approach nevertheless reached its performance ceiling. Therefore, in this paper, we investigate different enhancement techniques to boost the deep neural networks performance and provide a solution as BoostNet. The experiment is categorized into four different phases. We have selected ChampNet from benchmark deep learning models (EfficientNet: B0, MobileNet, ResNet18, VGG19). This phase helps to obtain the best model. In the second phase, The ChampNet evaluates with different resolution datasets. This phase helps to finalize the dataset resolution to enhance the performance of ChampNet. In the third phase, Champ-Net merges with image enhancement techniques, Contrast Limited Adaptive Histogram Equalization (CLAHE), High-frequency filtering (HEF), and Unsharp masking (UM). This phase helps to obtain Boost-Net with enriched performance. The last phase helps us to verify BoostNet results with Lightness Order Error. The presented research work fuses the image enhancement technique with ChampNet to generate BoostNet models. An assessment was performed on the Musculoskeletal Radiograph Bone Classification using classification schemes to demonstrate the proposed model's performance. The Classification accuracy of BoostNet was for the train a test dataset with and without enhancement techniques. The proposed model ChampNet + CLAHE, ChampNet + HEF, ChampNet + UM approach achieved 95.88%, 94.99%, and 94.18% accuracy, respectively. This experiment leads to a more accurate and efficient classification model. The main aim of this paper is to enhance techniques to boost the deep neural networks performance and provide a solution as BoostNet.

2D amorphous-MoO3−x@Ti3C2-MXene non-van der Waals heterostructures as anode materials for lithium-ion batteries

2D heterostructures offer a great opportunity in seeking high-performing energy storage materials; however, performance ceiling exists, limited by their van-der-Waals (vdW) interactions. Here, we explore a novel 2D, amorphous MoO3−x (aMoO3−x) on Ti3C2-MXene, non-vdW heterostructure via a facile synthesis route. Density functional theory computations suggest that the non-vdW heterostructure can strongly stabilize aMoO3−x while maintaining electrical conductivity at a high level. Facile 2D Li-ion diffusion can then be achieved in the restacked 2D non-vdW heterostructures due to the weakened interactions between two defective MoO3−x layers, leading to a capacitor-like interlayer diffusion reaching a large capacity of 426 C g−1 on the surface of the amorphous layer and a diffusion-controlled intralayer diffusion of 546 C g−1 within the amorphous layer. These characteristics optimize Li-ion storage kinetics while achieving full capacities of amorphous materials with high stability. This work might offer a feasible platform of 2D non-vdW heterostructures for boosting and understanding Li-ion storage performance.

Evaluating NISAR's cropland mapping algorithm over the conterminous United States using Sentinel-1 data

Accurate knowledge of the distribution, breadth and change in agricultural activity is important to food security and the related trade and policy mechanisms. Routine observations afforded by spaceborne Synthetic Aperture Radar (SAR) allows for high-fidelity monitoring of agricultural parameters at the field scale. Here we evaluate the approach to be used for generating NASA's upcoming NASA ISRO SAR (NISAR) mission's L-band cropland product using Sentinel-1C-band data. This study uses all ascending Sentinel-1A/B data collected over the conterminous United States in 2017 to compute the coefficient of variation (CV) at 150 m × 150 m resolution and evaluates the overall accuracy (OA) of CV-based crop/non-crop classifications at 100 one-by-one degree tiles. We calculate accuracies using two approaches: (a) using a literature-recommended constant CV threshold of 0.5 (CVthr_0.5) and (b) determining optimal CV thresholds for every tile using Youden's J statistic (YJS), CVthr_YJS. These accuracy comparisons are important for determining (1) the viability of using a computationally inexpensive and straightforward approach for cropland classification over large spatial scales/diverse land covers (i.e., can accuracies ≥80% be routinely achieved?), (2) how closely OA0.5 compares to the performance ceiling (OAYJS). This information will help determine whether approach (a) is appropriate and how much potential room of improvement there could be in modifying it. Results for OA0.5 and OAYJS are 81.5% and 86.8%, respectively. A breakdown by census geographic region, showed that OA0.5 (OAYJS) exceeded 80% (90%) in the South and Midwest, but was only 76.1% (73.5%) in the West. The improvement in OAYJS mainly stems from tiles with >40% crop prevalence having about 10% greater OA values. To better examine the potential of the approach for land cover classification, results of approach (b) were also stratified by crop. Approach (b) accurately detected most non-crop classes as non-crop (>80%), but with low OAYJS values for grasslands/pasture, especially in the West. CV values for crop were distinct from non-crop indicating that the approach is suitable for crop/non-crop classifications. Because results CV values have substantial overlap within crop/non-crop classes, indicating the approach is poorly suited for land cover classifications. We also detected a strong geographic dependence of CVthr_YJS: values ranged from about 0.2 at the coasts and gradually increase to about 0.6 in the Central United States, most often falling close to 0.3 and 0.5.

Predicting the Real-Valued Inter-Residue Distances for Proteins.

Predicting protein structure from the amino acid sequence has been a challenge with theoretical and practical significance in biophysics. Despite the recent progresses elicited by improved inter‐residue contact prediction, contact‐based structure prediction has gradually reached the performance ceiling. New methods have been proposed to predict the inter‐residue distance, but unanimously by simplifying the real‐valued distance prediction into a multiclass classification problem. Here, a lightweight regression‐based distance prediction method is shown, which adopts the generative adversarial network to capture the delicate geometric relationship between residue pairs and thus could predict the continuous, real‐valued inter‐residue distance rapidly and satisfactorily. The predicted residue distance map allows quick structure modeling by the CNS suite, and the constructed models approach the same level of quality as the other state‐of‐the‐art protein structure prediction methods when tested on CASP13 targets. Moreover, this method can be used directly for the structure prediction of membrane proteins without transfer learning.

Inverse design of an integrated-nanophotonics optical neural network

Artificial neural networks have dramatically improved the performance of many machine-learning applications such as image recognition and natural language processing. However, the electronic hardware implementations of the above-mentioned tasks are facing performance ceiling because Moore’s Law is slowing down. In this article, we propose an optical neural network architecture based on optical scattering units to implement deep learning tasks with fast speed, low power consumption and small footprint. The optical scattering units allow light to scatter back and forward within a small region and can be optimized through an inverse design method. The optical scattering units can implement high-precision stochastic matrix multiplication with mean squared error <10-4 and a mere 4 × 4 μm2 footprint. Furthermore, an optical neural network framework based on optical scattering units is constructed by introducing “Kernel Matrix”, which can achieve 97.1% accuracy on the classic image classification dataset MNIST.

FeatherCNN: Fast Inference Computation with TensorGEMM on ARM Architectures

Deep Learning is ubiquitous in a wide field of applications ranging from research to industry. In comparison to time-consuming iterative training of convolutional neural networks (CNNs), inference is a relatively lightweight operation making it amenable to execution on mobile devices. Nevertheless, lower latency and higher computation efficiency are crucial to allow for complex models and prolonged battery life. Addressing the aforementioned challenges, we propose FeatherCNN – a fast inference library for ARM CPUs – targeting the performance ceiling of mobile devices. FeatherCNN employs three key techniques: 1) A highly efficient TensorGEMM (generalized matrix multiplication) routine is applied to accelerate Winograd convolution on ARM CPUs, 2) General layer optimization based on custom high performance kernels improves both the computational efficiency and locality of memory access patterns for non-Winograd layers. 3) The framework design emphasizes joint layer-wise optimization using layer fusion to remove redundant calculations and memory movements. Performance evaluation reveals that FeatherCNN significantly outperforms state-of-the-art libraries. A forward propagation pass of VGG-16 on a 64-core ARM server is 48, 14, and 12 times faster than Caffe using OpenBLAS, Caffe2 using Eigen, and NNPACK, respectively. In addition, FeatherCNN is 3.19 times faster than the recently released TensorFlow Lite library on an iPhone 7 plus. In terms of GEMM performance, FeatherCNN achieves 14.8 and 39.0 percent higher performance than Apple's Accelerate framework on an iPhone 7 plus and Eigen on a Samsung Galaxy S8, respectively. The source code of FeatherCNN library is publicly available at https://github.com/tencent/feathercnn .

Robust and Low-Overhead Hybrid Beamforming Design With Imperfect Phase Shifters in Multi-User Millimeter Wave Systems

Phase shifters, often applied for analog beamforming in hybrid millimeter wave systems, are usually imperfect with random phase and gain errors brought by manufacture imperfections. Due to the uncertainty and nonreciprocity of random phase and gain errors, the digital precoder cannot eliminate the inter-user interference, which leads to system performance degradation. In our previous works [1], [2], we have analyzed the degradation of the achievable sum rate caused by imperfect phase shifters. The results show that the impact of random phase and gain errors is indeed very severe and they lead to performance ceiling. In this paper, we propose a novel channel estimation and Hybrid BeamForming (HBF) design method considering the structures both with and without switches. More specifically, we apply the discrete Fourier transform interpolation algorithm to estimate the downlink angle-of-departure of the strongest path, which avoids the exhaustive search of the narrow beam codebook and greatly saves the training overhead. Then, the downlink equivalent channel is estimated at the users and fed back for the digital precoder design, which guarantees the robustness against imperfect phase shifters. Finally, we derive a closed-form expression to make a better tradeoff between the estimation performance and training overhead. We show that our proposed method outperforms other state-of-the-art beam training methods with much less training overhead both theoretically and numerically.

Deep Channel-Correlation Network for Motor Imagery Decoding From the Same Limb

Motor imagery (MI) is an important brain-computer interface (BCI) paradigm, which can be applied without external stimulus. Imagining different joint movements from the same limb allows intuitive control of the outer devices. However, few researches focused on this field, and the decoding accuracy limited the applications for practical use. In this study, we aim to use deep learning methods to explore the ceiling of the decoding performance of three tasks: the resting state, the MI of right hand and right elbow. To represent the brain functional relationships, the correlation matrix that consists of correlation coefficients between electrodes (channels) was calculated as features. We proposed the Channel-Correlation Network to learn the overall representation among channels for classification. Ensemble learning was applied to integrate the output of multiple Channel-Correlation Networks. Our proposed method achieved the decoding accuracy of up to 87.03% in the 3-class scenario. The results demonstrated the effectiveness of deep learning method for decoding MI of different joints from the same limb and the potential of this fine paradigm to be applied in practice.

Interface modulation of twinned PtFe nanoplates branched 3D architecture for oxygen reduction catalysis

Highly-branched dendritic Pt-based nanocrystals possess great potential in catalyzing the oxygen reduction reaction (ORR), but encounter performance ceiling due to their poor thermal and electrochemical stability. Here, we present a novel PtFe nanodendrites (NDs) branched with two-dimensional (2D) twinned nanoplates rather than conventional 1D nanowires, which breaks the ORR performance ceiling of dendritic catalysts by inducing the unique Pt-skin configuration via rationally thermal treatment. By further hybridizing the Pt-skin PtFe NDs/C with amino-functionalized ionic liquids (ILs), we achieve an unprecedented mass activity of 3.15 A/mgPt at 0.9 V versus reversible hydrogen electrode (RHE) in the PtFe-based ORR electrocatalytic system. They also show excellent electrocatalytic durability for ORR with negligible activity decay and no apparent structural change after 20,000 cycles, in sharp contrast to the nanowires branched PtFe NDs counterpart. The remarkable catalytic performance is attributed to a combination of several structural features, including 2D morphology, twin boundary, partially ordered phase and strong coordination with amino group. This work highlights the significance of stabilizing electrocatalytic structures via morphology tuning, which thus enables further surface and interface modification for performance breakthrough in ORR electrocatalysis.

Facilitators and barriers to implement the family doctor contracting services in China: findings from a qualitative study

ObjectiveTo identify the facilitators and barriers to implement family doctor contracting services in China by using Consolidated Framework for Implementation Research (CFIR) to shed new light on establishing family doctor...

Measuring similarity and relatedness using multiple semantic relations in WordNet

Semantic similarity and relatedness computation has attracted an increasing amount of attention among researchers. The majority of previous studies, including edge-based and information content-based methods, rely on a single semantic relationship in WordNet such as the “is-a” relation. However, a performance ceiling may have been created by semantic unicity and inadequate calculation in solely “is-a” relation-based measurements, i.e., the computed results for some word pairs are too small and significantly deviate from human judgments. For this problem, we propose the following solutions: (1) We introduce the notion of the nearest common descendant to provide a supplement for commonalities between concepts according to genetics theory. (2) We design various targeted methods for different incomplete semantic relations. Therefore, various semantic relations can participate in similarity and relatedness computations in their most appropriate manners. (3) We utilize the cross-use of incomplete semantic relations similar-to and antonymy to solve the challenge of adjective and adverb similarity/relatedness measurements in WordNet. (4) We propose a targeted independent computation and largest contribution aggregation method to break through the performance ceiling of similarity/relatedness measurements based on single “is-a” relations. We conduct evaluations of our proposed model using seven extensively employed datasets. These evaluations indicate that our method significantly improves the performance of the existing methods based on single “is-a” relations. Their best Pearson coefficient with human judgments on both the MC30 and RG65 is increased to 0.9. With the development and enrichment of semantic relations in WordNet, our proposed model can be expected to have a more prominent role.

Spectral Efficiency of Full-Duplex Wireless Backhaul with Hardware Impaired Massive MIMO for Heterogeneous Cellular Networks

The paper analyzes the sum spectral efficiency (SE) for a heterogeneous cellular network (HetNet) which has the backhaul, provided with wireless full-duplex massive multiple-input multiple-out (MIMO) with hardware distortions. We derive approximate expressions to obtain the uplink/downlink sum SE of the backhaul. The analytic results have been shown to be exact when compared to Monte Carlo simulations. From the analysis, it is shown that the desired signal and the hardware distortion noise have the same order. The sum SE generally improves when the number of receive antennas increases but degrades when the hardware quality reduces. A sum SE performance ceiling is introduced by the hardware quality level.

3-D Wire Bondless Switching Cell Using Flip-Chip-Bonded Silicon Carbide Power Devices

This paper presents a three-dimensional (3-D) wire bondless power module using silicon carbide (SiC) power devices. Commercially available SiC power devices are designed for wire bonding. Wire bonds have an inherent parasitic inductance that limits high-frequency switching. This results in an underutilization of the full potential of SiC power devices, which have very low switching losses at high frequencies. Wire-bonded power modules run into a performance ceiling when it comes to ultrafast switching. This paper strives to provide a solution to this issue, which involves reconfiguring a commercially available bare die SiC power device into a flip-chip-capable device. A wire bondless SiC Schottky diode package was demonstrated and its performance was contrasted with a conventional wire-bonded package. A 24% reduction in the ON-state resistance was observed in the wire bondless package. As a next step, wire bondless SiC MOSFET packages were developed and tested in a half-bridge configuration in a highly integrated 3-D arrangement. This approach departs from the conventional concept of a power module—demonstrating a direct-bonded-copper-less and baseplate-less half-bridge switching cell. Double-pulse tests conducted on the cell showed >3× reduction in the parasitic inductance of the 3-D cell as compared with a conventional wire-bonded module.