Backward Propagation Research Articles

Recombination of N Atoms in a Manifold of Electronic States Simulated by Time-Reversed Nonadiabatic Photodissociation Dynamics of N2.

Following a single photon VUV absorption, the N2 molecule dissociates into distinct channels leading to N atoms of different reactivities. The optically accessible singlets are bound, and dissociation occurs through spin-orbit induced transfer to the triplets. There is a forest of coupled electronic states, and we here aim to trace a path along the nonadiabatic couplings toward a particular exit channel. To achieve this, we apply a time-reversed quantum dynamical approach that corresponds to a dissociation running back. It begins with an atom-atom relative motion in a particular product channel. Starting with a Gaussian wave packet at the dissociation region of N2 and propagating it backward in time, one can see the population transferring among the triplets due to a strong nonadiabatic interaction between these states. Simultaneously, the optically active singlets get populated because of spin-orbit coupling to the triplets. Thus, backward propagation traces the nonradiative association of nitrogen atoms.

Artificial neural network coupled experimental investigation of tribological behavior of additive manufactured PETG/OMMT nanocomposites

AbstractThe advent of additive manufacturing has brought a paradigm shift in many engineering applications involving polymer composites. The current study aims at integrating artificial neural network (ANN) with experimentation to assess the wear properties of glycol‐modified polyethylene terephthalate (PETG) composites reinforced with organically modified montmorillonite (OMMT) nanoclay (1%, 3%, and 5% weight percentages). The specimens are fabricated using fused deposition modeling technology. The proposed composites are compounded and developed using a single screw extruder with a diameter of 1.75 mm. As per ASTM G99 standard, the tribological characteristics were investigated using sliding wear with the dry pin on disc arrangement. The data points owing to the wear properties of PETG/OMMT nanocomposites are collected with the aid of the experimentation, which is then used to train the ANN model using Levenberg– Marquardt backward propagation algorithm. The experimental results show that adding OMMT nanoclay at a 3% weight percentage improves the wear properties of the composites compared to virgin PETG. A scanning electron microscope study of the wear mechanism reveals that adding OMMT nanoclay to the proposed composites results in mild wear as opposed to heavy wear in the case of virgin PETG. The ANN model developed to predict the tribological performance of PETG/OMMT facilitates better predictability with 99.6% accuracy, thereby reducing the requirement of expensive experimentation and analysis. The proposed composites may be applicable in the prosthetic, aerospace and automobile industries.

Analysis of Influence of the Seamount in the South China Sea on Sound Propagation

The influence of seamounts on sound propagation is very important, and it is also a research hotspot for hydroacoustic. This paper uses the BELLHOP ray model to analyze the influence of seamounts in the South China Sea on sound propagation, it is found that the blocking of seamounts has a greater impact on the spatial distribution of the transmission loss. When forward propagation, the distribution of the transmission loss is mainly influenced by seamounts, and its temporal variation is not obvious. When the sound source is relatively far away from the first seamount, the first seamount does not have a great influence on the distribution of the transmission loss. When backward propagation, the transmission loss increases rapidly after passing through the first seamount. Most of the sound energy is blocked due to the higher height of the first seamount. In addition, the impact of the undulating bottom on the front of seamounts on sound propagation should also not be ignored during backward propagation. The seamount determines the most important characteristics of the transmission loss distribution, while the undulating bottom has a certain degree of influence on the distribution characteristics of some specific regions. The above results have important applications for target detection and hydroacoustic communication in the area of seamounts.

Landing Reliability Assessment of Airdrop System Based on Vine-Bayesian Network

The landing phase of an airdrop process is prone to accidents, and thus, it is important to assess the landing reliability for an airdrop system. However, full field tests to assess the reliability are unacceptable due to their cost and the time required. As such, it is necessary to estimate the reliability in the design stage. To address this problem, a method based on vine-Bayesian Network (vine-BN) is proposed to assess the landing reliability by fusing multisource information. First, the network structure is determined by the relationship between data of simulation or ground tests and failure modes. Then, nodes are defined as random variables on [0, 1] based on the definition of the performance metric. Finally, the dependence between nodes is quantified by expert opinions. To illustrate the effectiveness of the method, a particular ground test or simulation is chosen to establish a network for a typical heavy cargo airdrop system (HCADS). Forward and backward propagation is carried out on the network. The forward analysis predicts the landing reliability in the design stage through multisource information fusion. Beta distribution is applied to fit the fusion result, so Bayesian inference is made to perform field test times decision-making. The backward analysis works to identify the key performance metrics related to landing reliability. The results and analysis manifest that vine-BN is feasible for fusing multisource information. Through the network, the reliability of the current design can be predicted effectively, and the field test times can be remarkably reduced. This method plays a crucial role in airdrop system design and reducing test time and labor.

Gradient distribution-aware INT8 training for neural networks

Recently, low bit-width quantization (e.g., INT8) has been commonly used in deep neural network inference acceleration, but fewer researchers have focused on low-precision training quantization techniques. Considering that the backward propagation in deep neural network training is more computationally intensive and has a heavier energy overhead than the inference process, the quantization of backward propagation is of great interest for the training of very large-scale neural networks as well as for low-power devices with online training requirements. However, the shape specificity and continuous variability of the gradient distribution make gradient quantization difficult, and many studies propose various complex quantization methods for the gradient to reduce training accuracy loss. In this paper, we propose two innovative techniques mainly for INT8 quantization training, including the Data-aware Dynamic Segmentation Quantization scheme to quantize various special gradient distributions and the Update Direction Periodic Search strategy to achieve lower quantization errors. Then, we build a distribution-aware INT8 quantization training framework based on these two methods and conduct experiments on various models and tasks. Experimental results show that our proposed INT8 quantization training method achieves a negligible loss in final training accuracy compared to the full-precision floating-point counterpart on different models, including ResNet, MobileNetV2, VGG, AlexNet, and LSTM. By replacing floating-point computing with 8-bit integer computing for network training, this INT8 quantization training framework provides the possibility of deploying online training directly on low-power devices in the future.

ANNbased prediction of the absorption characteristics of additive manufactured glycol‐modified polyethylene terephthalate nanocomposites reinforced with short‐carbon fibers and nanoclay fillers

AbstractIn this research work, an artificial neural network (ANN) based data driven prediction technique is proposed to evaluate the absorption characteristics of glycol‐modified polyethylene terephthalate (PETG) polymer nanocomposites reinforced with short carbon fibers (SCFs) and organically modified montmorillonite (OMMT) nanoclay (NC) fillers. The specimens are fabricated through additive manufacturing routine. The data points owing to the absorption of PETG nanocomposites are collected with the aid of the experimentation, which are then used to train the ANN model using Levenberg–Marquardt backward propagation algorithm. Different variants of PETG nanocomposites are prepared using various weight percentages of NC (1%, 3%, and 5%) and SCFs (5%, 10%, and 15%). The required filaments were prepared with the aid of compounding and extrusion processes. The specimens were printed using the fused deposition modeling technique. Further, the absorption behavior of these nanocomposites when exposed to distilled water, salt solution, and sugar solution has been experimentally studied following the ASTM D570 standard. The individual and combined effects of reinforcing OMMT‐NC and SCFs in the PETG matrix are also studied. The experimental results suggest that the absorption characteristics of PETG nanocomposites are significantly affected by the NC and SCFs reinforcements. The outcomes of this work will aid the design/material engineers in making a good trade‐off between the mechanical and absorption properties of the proposed PETG nanocomposites.

Over-the-Air Split Machine Learning in Wireless MIMO Networks

In split machine learning (ML), different partitions of a neural network (NN) are executed by different computing nodes, requiring a large amount of communication cost. As over-the-air computation (OAC) can efficiently implement all or part of the computation at the same time of communication, thus by substituting the wireless transmission in the traditional split ML framework with OAC, the communication load can be eased. In this paper, we propose to deploy split ML in a wireless multiple-input multiple-output (MIMO) communication network utilizing the intricate interplay between MIMO-based OAC and NN. The basic procedure of the OAC split ML system is first provided, and we show that the inter-layer connection in a NN of any size can be mathematically decomposed into a set of linear precoding and combining transformations over a MIMO channel carrying out multi-stream analog communication. The precoding and combining matrices which are regarded as trainable parameters, and the MIMO channel matrix, which are regarded as unknown (implicit) parameters, jointly serve as a fully connected layer of the NN. Most interestingly, the channel estimation procedure can be eliminated by exploiting the MIMO channel reciprocity of the forward and backward propagation, thus greatly saving the system costs and/or further improving its overall efficiency. The generalization of the proposed scheme to the conventional NNs is also introduced, i.e., the widely used convolutional NNs. We demonstrate its effectiveness under both the static and quasi-static memory channel conditions with comprehensive simulations.

Giant second harmonic generation in etch-less lithium niobate thin film

In this paper, we proposed and numerically demonstrated a giant enhancement up to in both fo108rward and backward propagation of the second harmonic generation by combining the high-quality factor cavities of the bound states in the continuum and the excellent nonlinear optical crystal of lithium niobate. The enhancement factor is defined as the ratio of the second harmonic signal generated by the structure (lithium niobate membrane with Si grating) divided by the signal generated by the lithium niobate membrane alone. Furthermore, a minimum interaction time of 350 ps is achieved despite the etching less lithium niobate membrane with a conversion efficiency of 4.77 × 10−6. The origin of the enhancements is linked to the excitation of a Fano-like shape symmetry-protected mode that is revealed by finite-difference time-domain simulations. The proposed platform opens the way to a new generation of efficient integrated optical sources compatible with nano-photonic devices for classical and quantum applications.

Bandgap merging and backward wave propagation in inertial amplification metamaterials

The aim of this study is to investigate the wave propagation characteristics of one-dimensional inertial amplification (IA) metamaterials with double resonators. The IA Angle is introduced by the IA mechanism as a critical parameter to modulate elastic waves. Double attenuation peaks in the bandgaps caused by coupling the IA mechanism and local resonance can be created in different ranges by adjusting the IA Angle. Moreover, a novel phenomenon of negative relative group velocity is observed in the region of double-negative effective properties, denoting negative energy flow. Numerical simulations using Gaussian pulse excitation are performed to validate the existence of a backward wave corresponding to the negative relative group velocity in the time domain. Furthermore, bandgap merging occurs in the region of double-zero effective properties, resulting in broadband wave attenuation. In particular, instead of using traditional metamaterials, two local resonator bandgaps can be merged to expand the bandgap range by at least 30% without changing the unit cell mass. The research results can be used to tune wave propagation in metamaterials, providing ideas for metamaterial design.

Lithium niobate thin film grating couplers optimized by particle swarm optimization and a neural network

Grating couplers (GCs) are a kind of critical device for integrated photonics, which connect on- and off-chip devices. In this paper, chirped GCs on Z-cut lithium niobate on insulator were designed and optimized using a backward propagation neural network (BPNN) combined with the particle swarm optimization (PSO) algorithm. The BPNN was proposed to predict the coupling efficiency (CE) of chirped GCs at hundreds of wavelengths simultaneously, which is 7400 times faster than finite difference time domain simulation. Furthermore, PSO was employed to search for the GC structures with high CE. The maximum CE that can be optimized through our trained network reaches 63% in 1550 nm. This work provides a fast and accurate method for designing efficient GCs at any central wavelength.

Multi-Depth Computer-Generated Hologram Based on Stochastic Gradient Descent Algorithm with Weighted Complex Loss Function and Masked Diffraction

In this paper, we propose a method to generate multi-depth phase-only holograms using stochastic gradient descent (SGD) algorithm with weighted complex loss function and masked multi-layer diffraction. The 3D scene can be represented by a combination of layers in different depths. In the wave propagation procedure of multiple layers in different depths, the complex amplitude of layers in different depths will gradually diffuse and produce occlusion at another layer. To solve this occlusion problem, a mask is used in the process of layers diffracting. Whether it is forward wave propagation or backward wave propagation of layers, the mask can reduce the occlusion problem between different layers. Otherwise, weighted complex loss function is implemented in the gradient descent optimization process, which analyzes the real part, the imaginary part, and the amplitude part of the focus region between the reconstructed images of the hologram and the target images. The weight parameter is used to adjust the ratio of the amplitude loss of the focus region in the whole loss function. The weight amplitude loss part in weighted complex loss function can decrease the interference of the focus region from the defocus region. The simulations and experiments have validated the effectiveness of the proposed method.

Bi-CapsNet: A Binary Capsule Network for EEG-based Emotion Recognition.

In recent years, deep learning has gained widespread attention in electroencephalogram (EEG)-based emotion recognition. However, deep learning methods are usually time-consuming with a large amount of memory usage, which obstructs their practical usage on resource-constrained devices. In this paper, we propose a binary capsule network (Bi-CapsNet) for EEG emotion recognition with low computational cost and memory usage. The Bi-CapsNet binarizes 32-bit weights and activations to 1 bit, and replaces floating-point operations with efficient bitwise operations. To address the issue of function discontinuity in backward propagation, we use a continuous function to approximate the binarization process. Two popular EEG emotion databases, namely, DEAP and DREAMER, are used for performance evaluation. In comparison to its full-precision counterpart, the Bi-CapsNet achieves a reduction on the computational cost and a reduction on the memory usage, while with only a 1% drop on the recognition accuracy. Compared to some state-of-the-art EEG emotion recognition methods, the proposed method obtains more competitive performance. In addition, the Bi-CapsNet is implemented on a mobile phone via an open-source binary inference framework named Bolt, and it achieves an ∼5× inference acceleration in comparison to its full-precision counterpart.

Dissipation-induced acoustic nonreciprocity

Linear acoustic transmission through a stationary elastic medium possesses Rayleigh reciprocal symmetry with respect to pressure produced by a source at point A(B) and measured at point B(A), pA(rB) = pB(rA). In ideal (inviscid) fluids, pressure plays the role of the scalar potential for velocity, v∼∇p. The Rayleigh reciprocity theorem is formulated for velocity potentials but not for their gradients. Any function of the absolute value of velocity, v(r), lacks reciprocal symmetry. In particular, acoustic intencity I = pv is not symmetric, IA(rB) ≠ IB(rA). Dynamics of ideal fluid is time-reversible, and lack of reciprocal symmetry is attributed to asymmetry in scattering. However, in a viscous environment dissipation breaks T symmetry making fluid dynamics irreversible. We report that acoustic transmission is not only irreversible but also becomes nonreciprocal due to different amount of dissipated energy for forward and backward propagation, provided that mirror symmetry (P symmetry) is broken. Since dissipation occurs within a narrow viscous layer at solid–fluid interface, the amount of nonreciprocity in transmission strongly depends on the quality of the surface of scatterers. Experiments performed with scatterers of different surface quality show how surface roughness affects the level of nonreciprocity. In a series of experiments with asymmetric phononic crystal, Tesla valve, and acoustic cavity, we demonstrate dissipation-induced nonreciprocity in transmission. [This work is supported by the NSF under EFRI Grant No. 1741677.]

Neuromorphic Computing Using Random Synaptic Feedback Weights for Error Backpropagation in NAND Flash Memory-Based Synaptic Devices

This work proposes utilizing separate synaptic string array for error backpropagation in NAND flash memory-based synaptic architecture with random synaptic feedback weight. To enable error backpropagation, forward and backward propagations are processed in separate synaptic devices in forward and backward synaptic arrays, respectively. In addition, synaptic weights in forward synaptic array are updated at each iteration, while those in backward synaptic array are fixed to reduce burden of peripheral circuits and power consumption. The optimal conductance response is investigated considering the linearity of the conductance response and the ratio of maximum and minimum currents. Reliability characteristics are verified by retention, endurance, and pass bias disturbance measurement results. Hardware-based neural networks with random synaptic weight achieve an inference accuracy of 95.41%, which is comparable to that of 95.58% obtained with transposed weight. Hardware-based neural network simulations demonstrate that the inference accuracy of the proposed on-chip learning scheme hardly decreases compared to that of the off-chip learning even with increasing device variation.

Modality-DTA: Multimodality Fusion Strategy for Drug-Target Affinity Prediction.

Prediction of the drug-target affinity (DTA) plays an important role in drug discovery. Existing deep learning methods for DTA prediction typically leverage a single modality, namely simplified molecular input line entry specification (SMILES) or amino acid sequence to learn representations. SMILES or amino acid sequences can be encoded into different modalities. Multimodality data provide different kinds of information, with complementary roles for DTA prediction. We propose Modality-DTA, a novel deep learning method for DTA prediction that leverages the multimodality of drugs and targets. A group of backward propagation neural networks is applied to ensure the completeness of the reconstruction process from the latent feature representation to original multimodality data. The tag between the drug and target is used to reduce the noise information in the latent representation from multimodality data. Experiments on three benchmark datasets show that our Modality-DTA outperforms existing methods in all metrics. Modality-DTA reduces the mean square error by 15.7% and improves the area under the precisionrecall curve by 12.74% in the Davis dataset. We further find that the drug modality Morgan fingerprint and the target modality generated by one-hot-encoding play the most significant roles. To the best of our knowledge, Modality-DTA is the first method to explore multimodality for DTA prediction.

Arterial wave dynamics preservation upon orthostatic stress: a modelling perspective.

Pressure-flow travelling waves are a key topic for understanding arterial haemodynamics. However, wave transmission and reflection processes induced by body posture changes have not been thoroughly explored yet. Current in vivo research has shown that the amount of wave reflection detected at a central level (ascending aorta, aortic arch) decreases during tilting to the upright position, despite the widely proved stiffening of the cardiovascular system. It is known that the arterial system is optimized when in the supine position, i.e. propagation of direct waves is enabled and reflected waves are trapped, protecting the heart; however, it is not known whether this is preserved with postural changes. To shed light on these aspects, we propose a multi-scale modelling approach to inquire into posture-induced arterial wave dynamics elicited by simulated head-up tilting. In spite of remarkable adaptation of the human vasculature following posture changes, our analysis shows that, upon tilting from supine to upright: (i) vessel lumens at arterial bifurcations remain well matched in the forward direction, (ii) wave reflection at central level is reduced due to the backward propagation of weakened pressure waves produced by cerebral autoregulation, and (iii) backward wave trapping is preserved.

T-PIM: An Energy-Efficient Processing-in-Memory Accelerator for End-to-End On-Device Training

Recently, on-device training has become crucial for the success of edge intelligence. However, frequent data movement between computing units and memory during training has been a major problem for battery-powered edge devices. Processing-in-memory (PIM) is a novel computing paradigm that merges computing logic into memory, which can address the data movement problem with excellent power efficiency. However, previous PIM accelerators cannot support the entire training process on chip due to its computing complexity. This article presents a PIM accelerator for end-to-end on-device training (T-PIM), the first PIM realization that enables end-to-end on-device training as well as high-speed inference. Its full-custom PIM macro contains 8T-SRAM cells to perform the energy-efficient in-cell AND operation and the bit-serial-based computation logic enables fully variable bit-precision for input data. The macro supports various data mapping methods and computational paths for both fully connected and convolutional layers, in order to handle the complex training process. An efficient tiling scheme is also proposed to enable T-PIM to compute any size of deep neural network with the implemented hardware. In addition, configurable arithmetic units in a forward propagation path make T-PIM handle power-of-two bit-precision for weight data, enabling a significant performance boost during inference. Finally, T-PIM efficiently handles sparsity in both operands by skipping the computation of zeros in the input data and by gating-off computing units when the weight data are zero. Finally, we fabricate the T-PIM chip in 28-nm CMOS technology, occupying a die area of 5.04 mm2, including five T-PIM cores. It dissipates 5.25–51.23 mW at 50–280 MHz operating frequency with 0.75–1.05-V supply voltage. We successfully demonstrate that T-PIM can run the end-to-end training of VGG16 model on the CIFAR10 and CIFAR100 datasets, achieving 0.13–161.08- and 0.25–7.59-TOPS/W power efficiency during inference and training, respectively. The result shows that T-PIM is <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$2.02\times $ </tex-math></inline-formula> more energy-efficient than the state-of-the-art PIM chip that only supports backward propagation, not a whole training. Furthermore, we conduct an architectural experiment using a cycle-level simulator based on actual measurement results, which suggests that the T-PIM architecture is scalable and its scaled-up version provides up to <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$203.26\times $ </tex-math></inline-formula> higher power efficiency than a comparable GPU.

Dynamic Graph Convolutional Recurrent Network for Traffic Prediction: Benchmark and Solution

Traffic prediction is the cornerstone of intelligent transportation system. Accurate traffic forecasting is essential for the applications of smart cities, i.e., intelligent traffic management and urban planning. Although various methods are proposed for spatio-temporal modeling, they ignore the dynamic characteristics of correlations among locations on road network. Meanwhile, most Recurrent Neural Network based works are not efficient enough due to their recurrent operations. Additionally, there is a severe lack of fair comparison among different methods on the same datasets. To address the above challenges, in this article, we propose a novel traffic prediction framework, named Dynamic Graph Convolutional Recurrent Network (DGCRN). In DGCRN, hyper-networks are designed to leverage and extract dynamic characteristics from node attributes, while the parameters of dynamic filters are generated at each time step. We filter the node embeddings and then use them to generate dynamic graph, which is integrated with pre-defined static graph. As far as we know, we are first to employ a generation method to model fine topology of dynamic graph at each time step. Furthermore, to enhance efficiency and performance, we employ a training strategy for DGCRN by restricting the iteration number of decoder during forward and backward propagation. Finally, a reproducible standardized benchmark and a brand new representative traffic dataset are opened for fair comparison and further research. Extensive experiments on three datasets demonstrate that our model outperforms 15 baselines consistently. Source codes are available at https://github.com/tsinghua-fib-lab/Traffic-Benchmark .

Prediction of Angular Distortion in Gas Metal Arc Welding of Structural Steel Plates Using Artificial Neural Networks

The manufacturing of structures ranging from bridges and machinery to all types of seaborne vehicles to nuclear reactors and space rockets has made considerable use of arc welding technologies. This is as a result of benefits including increased joint efficiency, air and water tightness, no thickness restriction (0.6 to 25 mm), decreased fabrication time and cost, etc. when compared to alternative fabrication methods. Gas metal arc welding (GMAW) is a frequently used welding technology in industries due to its inherent benefits, including deeper penetration, a smooth bead, etc. Local heating and cooling that takes place during the multi-pass welding process causes complicated stresses to develop at the weld zone, which ultimately causes angular distortion in the weldment. Angular distortion is a major flaw that affects the weld’s properties as well as the cracking and misalignment of the welded joints. The issue of angular distortion can be successfully solved by predicting it in relation to certain GMAW process variables. A neural network model was created in this research to predict angular distortion. A fractional factorial approach with 125 runs was used to conduct the exploratory experiments. A neural network model with feed forward and backward propagation was developed using the experimental data. To train the neural network model, the Levenberg–Marquardt method was utilised. The results indicate that the model based on network 4-9-3 is more effective in forecasting angular distortion with time gaps between two, three, and four passes than the other three networks (4-2-3, 4-4-3, 797 and 4-8-3). Prediction accuracy is more than 95 percent. The neural network model developed in this study can be used to manage the welding cycle in structural steel weld plates to achieve the best possible weld quality with the least amount of angular distortion.

A bidirectional DNN partition mechanism for efficient pipeline parallel training in cloud

Recently, deep neural networks (DNNs) have shown great promise in many fields while their parameter sizes are rapidly expanding. To break through the computation and memory limitation of a single machine, pipeline model parallelism is proposed for large-scale DNN training by fully utilizing the computation and storage power of the distributed cluster. Cloud data centers can also provide sufficient computing, storage and bandwidth resources. However, most existing approaches apply layer-wise partitioning, which is difficult to obtain an even model partition result because of the large computational overhead discrepancy between DNN layers, resulting in degraded efficiency. To tackle this issue, we propose “Bi-Partition”, a novel partitioning method based on bidirectional partitioning for forward propagation (FP) and backward propagation (BP), which improves the efficiency of the pipeline model parallelism system. By deliberated designing distinct cut positions for FP and BP of DNN training, workers in the pipeline get nearly equal computational loads, and the balanced pipeline fully utilizes the computing resources. Experiments on various DNN models and datasets validate the efficiency of our mechanism, e.g., the training efficiency achieving up to 1.9times faster than the state-of-the-art method PipeDream.