A Spiking Implementation of Backpropagation using Gated Synfire Chains
Explore the source record for details and available documents.
SEARCH · Engineering Papers
Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.
Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.
Explore the source record for details and available documents.
As an important class of spiking neural networks (SNNs), recurrent spiking neural networks (RSNNs) possess great computational power and have been widely used for processing sequential data like audio and text. However, most RSNNs suffer from two problems. 1. Due to the lack of architectural guidance, random recurrent connectivity is often adopted, which does not guarantee good performance. 2. Training of RSNNs is in general challenging, bottlenecking achievable model accuracy. To address these problems, we propose a new type of RSNNs called Skip-Connected Self-Recurrent SNNs (ScSr-SNNs). Recurrence in ScSr-SNNs is introduced in a stereotyped manner by adding self-recurrent connections to spiking neurons. The SNNs with self-recurrent connections can realize recurrent behaviors similar to those of more complex RSNNs while the error gradients can be more straightforwardly calculated due to the mostly feedforward nature of the network. The network dynamics is enriched by skip connections between nonadjacent layers. Moreover, we propose a new backpropagation (BP) method called backpropagated intrinsic plasticity (BIP) to further boost the performance of ScSr-SNNs by training intrinsic model parameters. Unlike standard intrinsic plasticity rules that adjust the neuron's intrinsic parameters according to neuronal activity, the proposed BIP method optimizes intrinsic parameters based on the backpropagated error gradient of a well-defined global loss function in addition to synaptic weight training. Here, based on challenging speech, neuromorphic speech, and neuromorphic image datasets, the proposed ScSr-SNNs can boost performance by up to 2.85% compared with other types of RSNNs trained by state-of-the-art BP methods.
Four important generic issues are identified and addressed in some depth in this thesis as part of the development of an adaptive neural network based control system for an experimental free flying space robot prototype. The first issue concerns the importance of true system level design of the control system. A new hybrid strategy is developed here, in depth, for the beneficial integration of neural networks into the total control system. A second important issue in neural network control concerns incorporating a priori knowledge into the neural network. In many applications, it is possible to get a reasonably accurate controller using conventional means. If this prior information is used purposefully to provide a starting point for the optimizing capabilities of the neural network, it can provide much faster initial learning. In a step towards addressing this issue, a new generic Fully Connected Architecture (FCA) is developed for use with backpropagation. A third issue is that neural networks are commonly trained using a gradient based optimization method such as backpropagation; but many real world systems have Discrete Valued Functions (DVFs) that do not permit gradient based optimization. One example is the on-off thrusters that are common on spacecraft. A new technique is developed here that now extends backpropagation learning for use with DVFs. The fourth issue is that the speed of adaptation is often a limiting factor in the implementation of a neural network control system. This issue has been strongly resolved in the research by drawing on the above new contributions.
This paper proposes a new method that combines checkpointing methods with error-controlled lossy compression for large-scale high-performance full-waveform inversion (FWI), an inverse problem commonly used in geophysical exploration. This combination can significantly reduce data movement, allowing a reduction in run time as well as peak memory. In the exascale computing era, frequent data transfer (e.g., memory bandwidth, PCIe bandwidth for GPUs, or network) is the performance bottleneck rather than the peak FLOPS of the processing unit. Like many other adjoint-based optimization problems, FWI is costly in terms of the number of floating-point operations, large memory footprint during backpropagation, and data transfer overheads. Past work for adjoint methods has developed checkpointing methods that reduce the peak memory requirements during backpropagation at the cost of additional floating-point computations. Combining this traditional checkpointing with error-controlled lossy compression, we explore the three-way tradeoff between memory, precision, and time to solution. We investigate how approximation errors introduced by lossy compression of the forward solution impact the objective function gradient and final inverted solution. Empirical results from these numerical experiments indicate that high lossy-compression rates (compression factors ranging up to 100) have a relatively minor impact on convergence rates and the quality of the final solution.
The effect of noise on the learning performance of the backpropagation algorithm is analyzed. A selective sampling of the training set is proposed to maximize the learning of control laws by backpropagation, when the data have been corrupted by noise. The training scheme is applied to the nonlinear control of a cart-pole system in the presence of noise. The neural computation provides the neurocontroller with good noise-filtering properties. In the presence of plant noise, the neurocontroller is found to be more stable than the teacher. A novel perspective on the application of neural network technology to control engineering is presented.
Conventional computing schemes have long been used to analyze problems in electromagnetics (EM). The vast majority of EM applications require computationally intensive algorithms involving numerical integration and solutions to large systems of equations. The feasibility of using neural network computing algorithms for antenna analysis is investigated. The ultimate goal is to use a trained neural network algorithm to reduce the computational demands of existing reflector surface error compensation techniques. Neural networks are computational algorithms based on neurobiological systems. Neural nets consist of massively parallel interconnected nonlinear computational elements. They are often employed in pattern recognition and image processing problems. Recently, neural network analysis has been applied in the electromagnetics area for the design of frequency selective surfaces and beam forming networks. The backpropagation training algorithm was employed to simulate classical antenna array synthesis techniques. The Woodward-Lawson (W-L) and Dolph-Chebyshev (D-C) array pattern synthesis techniques were used to train the neural network. The inputs to the network were samples of the desired synthesis pattern. The outputs are the array element excitations required to synthesize the desired pattern. Once trained, the network is used to simulate the W-L or D-C techniques. Various sector patterns and cosecant-type patterns (27 total) generated using W-L synthesis were used to train the network. Desired pattern samples were then fed to the neural network. The outputs of the network were the simulated W-L excitations. A 20 element linear array was used. There were 41 input pattern samples with 40 output excitations (20 real parts, 20 imaginary). A comparison between the simulated and actual W-L techniques is shown for a triangular-shaped pattern. Dolph-Chebyshev is a different class of synthesis technique in that D-C is used for side lobe control as opposed to pattern shaping. The interesting thing about D-C synthesis is that the side lobes have the same amplitude. Five-element arrays were used. Again, 41 pattern samples were used for the input. Nine actual D-C patterns ranging from -10 dB to -30 dB side lobe levels were used to train the network. A comparison between simulated and actual D-C techniques for a pattern with -22 dB side lobe level is shown. The goal for this research was to evaluate the performance of neural network computing with antennas. Future applications will employ the backpropagation training algorithm to drastically reduce the computational complexity involved in performing EM compensation for surface errors in large space reflector antennas.
In this paper we discuss a neural network architecture (the Probabilistic Neural Net or the PNN) that, to the best of our knowledge, has not previously been applied to remotely sensed data. The PNN is a supervised non-parametric classification algorithm as opposed to the Gaussian maximum likelihood classifier (GMLC). The PNN works by fitting a Gaussian kernel to each training point. The width of the Gaussian is controlled by a tuning parameter called the window width. If very small widths are used, the method is equivalent to the nearest neighbor method. For large windows, the PNN behaves like the GMLC. The basic implementation of the PNN requires no training time at all. In this respect it is far better than the commonly used backpropagation neural network which can be shown to take O(N6) time for training where N is the dimensionality of the input vector. In addition the PNN can be implemented in a feed forward mode in hardware. The disadvantage of the PNN is that it requires all the training data to be stored. Some solutions to this problem are discussed in the paper. Finally, we discuss the accuracy of the PNN with respect to the GMLC and the backpropagation neural network (BPNN). The PNN is shown to be better than GMLC and not as good as the BPNN with regards to classification accuracy.
As an initial step in the evaluation of networks, a feedforward architecture is trained to approximate experimental data by the backpropagation algorithm. Several drawbacks were detected and an alternative learning algorithm was then developed to partially address the drawbacks. This noniterative algorithm has a number of advantages over the backpropagation method and is easily implemented on existing hardware.
A new neural network architecture is proposed and applied in classification of remote sensing/geographic data from multiple sources. The new architecture is called the parallel consensual neural network and its relation to hierarchical and ensemble neural networks is discussed. The parallel consensual neural network architecture is based on statistical consensus theory. The input data are transformed several times and the different transformed data are applied as if they were independent inputs and are classified using stage neural networks. Finally, the outputs from the stage networks are then weighted and combined to make a decision. Experimental results based on remote sensing data and geographic data are given. The performance of the consensual neural network architecture is compared to that of a two-layer (one hidden layer) conjugate-gradient backpropagation neural network. The results with the proposed neural network architecture compare favorably in terms of classification accuracy to the backpropagation method.
Described herein is a method and system for training nonlinear adaptive filters (or neural networks) which have embedded memory. Such memory can arise in a multi-layer finite impulse response (FIR) architecture, or an infinite impulse response (IIR) architecture. We focus on filter architectures with separate linear dynamic components and static nonlinear components. Such filters can be structured so as to restrict their degrees of computational freedom based on a priori knowledge about the dynamic operation to be emulated. The method is detailed for an FIR architecture which consists of linear FIR filters together with nonlinear generalized single layer subnets. For the IIR case, we extend the methodology to a general nonlinear architecture which uses feedback. For these dynamic architectures, we describe how one can apply optimization techniques which make updates closer to the Newton direction than those of a steepest descent method, such as backpropagation. We detail a novel adaptive modified Gauss-Newton optimization technique, which uses an adaptive learning rate to determine both the magnitude and direction of update steps. For a wide range of adaptive filtering applications, the new training algorithm converges faster and to a smaller value of cost than both steepest-descent methods such as backpropagation-through-time, and standard quasi-Newton methods. We apply the algorithm to modeling the inverse of a nonlinear dynamic tracking system 5, as well as a nonlinear amplifier 6.
The VIVACE Converter consists of cylindrical oscillators in tandem subjected to transverse flow-induced oscillations (FIOs) that can be improved by varying the system parameters for a given in-flow velocity: damping, stiffness, and in-flow center-to-center spacing. Compared to a single isolated cylinder, tandem cylinders can harness more hydrokinetic energy due to synergy in FIO. Experimental and numerical methods have been utilized to analyze the FIO and energy harnessing of VIVACE. A surrogate-based model of two tandem cylinders is developed to predict the power harvesting and corresponding efficiency by introducing a backpropagation neural network. It is then utilized to reduce excessive experimental or computational testing. The effects of spacing, damping, and stiffness on harvested power and efficiency of the established prediction-model are analyzed. At each selected flow velocity, optimization results of power harvesting using the prediction-model are calculated under different combinations of damping and stiffness. The main conclusions are: (1) The surrogate model, built on extensive experimental data for tandem cylinders, can predict the cylinder oscillatory response accurately. (2) Increasing the damping ratio range from 0–0.24 to 0–0.30 is beneficial for improving power efficiency, but has no significant effect on power harvesting. (3) In galloping, a spacing ratio of 1.57 has the highest optimal harnessed power and efficiency compared with other spacing values. (4) Two tandem cylinders can harness 2.01–4.67 times the optimal power of an isolated cylinder. In addition, the former can achieve 1.46–4.01 times the efficiency of the latter. (5) The surrogate model is an efficient predictive tool defining parameters of the Converter for improved energy acquisition.
Abstract In order to eliminate the influence of scene illumination on the evaluation of the color difference of dyed fabrics, this paper proposes a dyed fabric illumination correction algorithm based on the extreme learning machine (ELM) with grey wolf optimizer (GWO)‐optimized ant lion optimizer (ALO). Firstly, the Grey Edge framework is used to extract the features of the dyed fabric image as the input vector. Then, to improve the optimization ability of the ALO algorithm, the GWO algorithm is used to provide a set of optimized initial populations to the ALO algorithm, and then the improved ALO algorithm is used to optimize the parameters of the ELM. Finally, the proposed GWO‐ALO‐ELM algorithm is used to correct the illumination of the dyed fabric, and restore the graphics to the effect display under standard illumination through the diagonal reduction model. Compared with the experimental results of GWO‐ELM, ALO‐ELM, backpropagation (BP), ELM, random vector function link (RVFL), and other algorithms, it can be seen that the GWO‐ALO‐ELM algorithm proposed in this paper has good predictive value and quasi‐bias effect, and good stability.
Quantitatively connecting properties of parton distribution functions (PDFs, or parton densities) to the theoretical assumptions made within the QCD analyses which produce them has been a longstanding problem in HEP phenomenology. To confront this challenge, we introduce an ML-based explainability framework, XAI4PDF, to classify PDFs by parton flavor or underlying theoretical model using ResNet-like neural networks (NNs). By leveraging the differentiable nature of ResNet models, this approach deploys guided backpropagation to dissect relevant features of fitted PDFs, identifying x-dependent signatures of PDFs important to the ML model classifications. By applying our framework, we are able to sort PDFs according to the analysis which produced them while constructing quantitative, human-readable maps locating the x regions most affected by the internal theory assumptions going into each analysis. This technique expands the toolkit available to PDF analysis and adjacent particle phenomenology while pointing to promising generalizations.
Using persistent homology to guide optimization has emerged as a novel application of topological data analysis. Existing methods treat persistence calculation as a black box and backpropagate gradients only onto the simplices involved in particular pairs. We show how the cycles and chains used in the persistence calculation can be used to prescribe gradients to larger subsets of the domain. In particular, we show that in a special case, which serves as a building block for general losses, the problem can be solved exactly in linear time. This relies on another contribution of this paper, which eliminates the need to examine a factorial number of permutations of simplices with the same value. Here, we present empirical experiments that show the practical benefits of our algorithm: the number of steps required for the optimization is reduced by an order of magnitude.
Radiation therapy plays an essential role in the treatment of cancer. In radiation therapy, the ideal radiation doses are delivered to the observed tumor while not affecting neighboring normal tissues. In three-dimensional computed tomography (3D-CT) scans, the contours of tumors and organs-at-risk (OARs) are often manually delineated by radiologists. The task is complicated and time-consuming, and the manually delineated results will be variable from different radiologists. We propose a semi-supervised contour detection algorithm, which firstly uses a few points of region of interest (ROI) as an approximate initialization. Then the data sequences are achieved by the closed polygonal line (CPL) algorithm, where the data sequences consist of the ordered projection indexes and the corresponding initial points. Finally, the smooth lung contour can be obtained, when the data sequences are trained by the backpropagation neural network model (BNNM). We use the private clinical dataset and the public Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) dataset to measure the accuracy of the presented method, respectively. To the private dataset, experimental results on the initial points which are as low as 15% of the manually delineated points show that the Dice coefficient reaches up to 0.95 and the global error is as low as 1.47 × 10{sup −2}. The performance of the proposed algorithm is also better than the cubic spline interpolation (CSI) algorithm. While on the public LIDC-IDRI dataset, our method achieves superior segmentation performance with average Dice of 0.83.
This work proposes the use of orthogonal bipolar vectors (OBV) as targets for multilayer perceptron (MLP) Artificial neural networks pertaining to the pattern recognition field. Implications involving this proposal are supported by a proper mathematical discussion. This study shows that the use of OBV as targets improves the class identification performance of MLP networks for pattern recognition with more than two classes when compared to the widely used conventional target bipolar vector (CBV). In addition to the overall performance improvement, it is possible to obtain high recognition rates with fewer training epochs, while keeping a feasible generalization with other data. This behavior is related to the Euclidean distance of the points in the output space, which is much higher in the case of OBVs. Therefore, with a greater distance between the points in the output space, the MLP network becomes less prone to associate trained outputs with incorrect targets. The mathematical discussion in this study is based on the backpropagation training algorithm. In addition to the mathematical discussion, results from experiments with real data are presented. These results follow those obtained by mathematical derivations. Real data sets used in the experiments are available from: (a) the Semeion Handwritten Digit of Machine Learning Repository, international repository; (b) Iris Image Database from the Chinese Academy of Sciences—CASIA; (c) Australian Sign Language, signs of Machine Learning Repository, international repository. Experiments were conducted to measure the mean Euclidean distance between the MLP network output subjected to a given input pattern and outputs related to other pattern classifications. Experimental results show that MLP networks trained with OBVs were able to ensure a Euclidean distance between the output and incorrect targets greater than MLP networks trained with CBVs (by up to 4 times for handwritten, by up to 12 times for iris images and by up to 10 times for Auslan signs language). As such, the adoption of OBVs as target vectors for MLPs reduces the computational effort for training the network.
Multimodal luminescent materials, while promising for anti-counterfeiting, often lack dynamic time-dependent responses and controllable spatial distribution, limiting their encryption capabilities in the spatiotemporal dimension. Here, this work presents a coordinated control strategy based on entropy and defect engineering, and uses a backpropagation (BP) neural network for material screening to successfully prepare spinel Mg 0.8 (Fe 0.04 Co 0.04 Ni 0.04 Cu 0.04 Zn 0.04 )Cr 2 O 4 (MgA 5 CO) phosphors with time-dependent dynamic luminescence behavior. This phosphor simultaneously activated the d-d transition luminescence (∼618 nm) derived from Co 2+ /Cr 3+ and the defect luminescence (∼398 nm) related to zinc vacancies (V Zn ) in a single-phase solid solution. The phosphor exhibits a time-dependent color evolution from pink to purple under fixed-wavelength excitation, due to the different excited-state dynamics and decay lifetimes associated with the d-d transition and defect luminescence. Structural characterization and spectral analysis confirmed the existence of V Zn and its significant role in defect luminescence process. The fluorescent and dynamic luminescent properties of entropy-based spinel oxide enable its use in advanced anti-counterfeiting applications like fingerprint recognition and color-changing dedicated anti-counterfeiting mark, showing promise in high-end and time-dynamic anti-counterfeiting fields. This research not only developed a new type of fluorescent dynamic anti-counterfeiting material, but also provided a new idea for constructing advanced optical functional materials with multiple luminescence mechanisms.
We present a differentiable predictive control (DPC) methodology for learning constrained control laws for unknown nonlinear systems. DPC poses an approximate solution to multiparametric programming problems emerging from explicit nonlinear model predictive control (MPC). Contrary to approximate MPC, DPC does not require supervision by an expert controller. Instead, a system dynamics model is learned from a small dataset of recorded observations of the perturbed system's dynamics and the control law is optimized offline by interaction with the learned system model. The DPC method is based on two sequential steps, i) system identification using a constrained neural state-space model, and ii) optimization of an explicit control law parametrized by another neural network in closed-loop simulation with the identified neural state-space model. The combination of a differentiable closed-loop system and penalty methods for constraint handling of system outputs and inputs allows us to optimize the control law's parameters directly by backpropagating economic MPC loss through the learned system model. By incorporating domain knowledge and leveraging established techniques from optimal control, our method leverages deep neural networks as nonlinear function approximators for system identification and control while avoiding concomitant costs of intractably large datasets, and computationally expensive over-parametrized models. The scalability, data efficiency, and constrained optimal control capability of the proposed DPC method are demonstrated in simulation using a multi-zone building emulator.