NeMo: a platform for neural modelling of spiking neurons using GPUs

Simulating spiking neural networks is of great interest to scientists wanting to model the functioning of the brain. However, large-scale models are expensive to simulate due to the number and interconnectedness of neurons in the brain. Furthermore, where such simulations are used in an embodied setting, the simulation must be real-time in order to be useful. In this paper we present NeMo, a platform for such simulations which achieves high performance through the use of highly parallel commodity hardware in the form of graphics processing units (GPUs). NeMo makes use of the Izhikevich neuron model which provides a range of realistic spiking dynamics while being computationally efficient. Our GPU kernel can deliver up to 400 million spikes per second. This corresponds to a real-time simulation of around 40 000 neurons under biologically plausible conditions with 1000 synapses per neuron and a mean firing rate of 10 Hz.

Mind wandering away from pain dynamically engages antinociceptive and default mode brain networks

Human minds often wander away from their immediate sensory environment. It remains unknown whether such mind wandering is unsystematic or whether it lawfully relates to an individual’s tendency to attend to salient stimuli such as pain and their associated brain structure/function. Studies of pain–cognition interactions typically examine explicit manipulation of attention rather than spontaneous mind wandering. Here we sought to better represent natural fluctuations in pain in daily life, so we assessed behavioral and neural aspects of spontaneous disengagement of attention from pain. We found that an individual’s tendency to attend to pain related to the disruptive effect of pain on his or her cognitive task performance. Next, we linked behavioral findings to neural networks with strikingly convergent evidence from functional magnetic resonance imaging during pain coupled with thought probes of mind wandering, dynamic resting state activity fluctuations, and diffusion MRI. We found that (i) pain-induced default mode network (DMN) deactivations were attenuated during mind wandering away from pain; (ii) functional connectivity fluctuations between the DMN and periaqueductal gray (PAG) dynamically tracked spontaneous attention away from pain; and (iii) across individuals, stronger PAG–DMN structural connectivity and more dynamic resting state PAG–DMN functional connectivity were associated with the tendency to mind wander away from pain. These data demonstrate that individual tendencies to mind wander away from pain, in the absence of explicit manipulation, are subserved by functional and structural connectivity within and between default mode and antinociceptive descending modulation networks.

Smart-phone based electrocardiogram wavelet decomposition and neural network classification

This paper discusses ECG classification after parametrizing the ECG waveforms in the wavelet domain. The aim of the work is to develop an accurate classification algorithm that can be used to diagnose cardiac beat abnormalities detected using a mobile platform such as smart-phones. Continuous time recurrent neural network classifiers are considered for this task. Records from the European ST-T Database are decomposed in the wavelet domain using discrete wavelet transform (DWT) filter banks and the resulting DWT coefficients are filtered and used as inputs for training the neural network classifier. Advantages of the proposed methodology are the reduced memory requirement for the signals which is of relevance to mobile applications as well as an improvement in the ability of the neural network in its generalization ability due to the more parsimonious representation of the signal to its inputs.

B-spline neural network based single-carrier frequency domain equalisation for Hammerstein channels

A practical single-carrier (SC) block transmission with frequency domain equalisation (FDE) system can generally be modelled by the Hammerstein system that includes the nonlinear distortion effects of the high power amplifier (HPA) at transmitter. For such Hammerstein channels, the standard SC-FDE scheme no longer works. We propose a novel Bspline neural network based nonlinear SC-FDE scheme for Hammerstein channels. In particular, we model the nonlinear HPA, which represents the complex-valued static nonlinearity of the Hammerstein channel, by two real-valued B-spline neural networks, one for modelling the nonlinear amplitude response of the HPA and the other for the nonlinear phase response of the HPA. We then develop an efficient alternating least squares algorithm for estimating the parameters of the Hammerstein channel, including the channel impulse response coefficients and the parameters of the two B-spline models. Moreover, we also use another real-valued B-spline neural network to model the inversion of the HPA’s nonlinear amplitude response, and the parameters of this inverting B-spline model can be estimated using the standard least squares algorithm based on the pseudo training data obtained as a byproduct of the Hammerstein channel identification. Equalisation of the SC Hammerstein channel can then be accomplished by the usual one-tap linear equalisation in frequency domain as well as the inverse Bspline neural network model obtained in time domain. The effectiveness of our nonlinear SC-FDE scheme for Hammerstein channels is demonstrated in a simulation study.

Cross-situational learning of object-word mapping using Neural Modeling Fields

The issue of how children learn the meaning of words is fundamental to developmental psychology. The recent attempts to develop or evolve efficient communication protocols among interacting robots or Virtual agents have brought that issue to a central place in more applied research fields, such as computational linguistics and neural networks, as well. An attractive approach to learning an object-word mapping is the so-called cross-situational learning. This learning scenario is based on the intuitive notion that a learner can determine the meaning of a word by finding something in common across all observed uses of that word. Here we show how the deterministic Neural Modeling Fields (NMF) categorization mechanism can be used by the learner as an efficient algorithm to infer the correct object-word mapping. To achieve that we first reduce the original on-line learning problem to a batch learning problem where the inputs to the NMF mechanism are all possible object-word associations that Could be inferred from the cross-situational learning scenario. Since many of those associations are incorrect, they are considered as clutter or noise and discarded automatically by a clutter detector model included in our NMF implementation. With these two key ingredients - batch learning and clutter detection - the NMF mechanism was capable to infer perfectly the correct object-word mapping. (C) 2009 Elsevier Ltd. All rights reserved.

How language can help discrimination in the Neural Modelling Fields framework

The relationship between thought and language and, in particular, the issue of whether and how language influences thought is still a matter of fierce debate. Here we consider a discrimination task scenario to study language acquisition in which an agent receives linguistic input from an external teacher, in addition to sensory stimuli from the objects that exemplify the overlapping categories that make up the environment. Sensory and linguistic input signals are fused using the Neural Modelling Fields (NMF) categorization algorithm. We find that the agent with language is capable of differentiating object features that it could not distinguish without language. In this sense, the linguistic stimuli prompt the agent to redefine and refine the discrimination capacity of its sensory channels. (C) 2007 Elsevier Ltd. All rights reserved.

A multi-GPU algorithm for large-scale neuronal networks

Large-scale simulations of parts of the brain using detailed neuronal models to improve our understanding of brain functions are becoming a reality with the usage of supercomputers and large clusters. However, the high acquisition and maintenance cost of these computers, including the physical space, air conditioning, and electrical power, limits the number of simulations of this kind that scientists can perform. Modern commodity graphical cards, based on the CUDA platform, contain graphical processing units (GPUs) composed of hundreds of processors that can simultaneously execute thousands of threads and thus constitute a low-cost solution for many high-performance computing applications. In this work, we present a CUDA algorithm that enables the execution, on multiple GPUs, of simulations of large-scale networks composed of biologically realistic Hodgkin-Huxley neurons. The algorithm represents each neuron as a CUDA thread, which solves the set of coupled differential equations that model each neuron. Communication among neurons located in different GPUs is coordinated by the CPU. We obtained speedups of 40 for the simulation of 200k neurons that received random external input and speedups of 9 for a network with 200k neurons and 20M neuronal connections, in a single computer with two graphic boards with two GPUs each, when compared with a modern quad-core CPU. Copyright (C) 2010 John Wiley & Sons, Ltd.

Predicting worsted spinning performance with an artificial neural network model

For a given fiber spun to pre-determined yarn specifications, the spinning performance of the yarn usually varies from mill to mill. For this reason, it is necessary to develop an empirical model that can encompass all known processing variables that exist in different spinning mills, and then generalize this information and be able to accurately predict yarn quality for an individual mill. This paper reports a method for predicting worsted spinning performance with an artificial neural network (ANN) trained with backpropagation. The applicability of artificial neural networks for predicting spinning performance is first evaluated against a well established prediction and benchmarking tool (Sirolan YarnspecTM). The ANN is then subsequently trained with commercial mill data to assess the feasibility of the method as a mill-specific performance prediction tool. Incorporating mill-specific data results in an improved fit to the commercial mill data set, suggesting that the proposed method has the ability to predict the spinning performance of a specific mill accurately.

Predicting the pilling propensity of fabrics through artificial neural network modeling

Fabric pilling is affected by many interacting factors. This study uses artificial neural networks to model the multi-linear relationships between fiber, yarn and fabric properties and their effect on the pilling propensity of pure wool knitted fabrics. This tool shall enable the user to gauge the expected pilling performance of a fabric from a number of given inputs. It will also provide a means of improving current products by offering alternative material specification and/or selection. In addition to having the capability to predict pilling performance, the model will allow for clarification of major fiber, yarn and fabric attributes affecting fabric pilling.

Knowledge extraction from a mixed transfer function artificial neural network

One of the main problems with Artificial Neural Networks (ANNs) is that their results are not intuitively clear. For example, commonly used hidden neurons with sigmoid activation function can approximate any continuous function, including linear functions, but the coefficients (weights) of this approximation are rather meaningless. To address this problem, current paper presents a novel kind of a neural network that uses transfer functions of various complexities in contrast to mono-transfer functions used in sigmoid and hyperbolic tangent networks. The presence of transfer functions of various complexities in a Mixed Transfer Functions Artificial Neural Network (MTFANN) allow easy conversion of the full model into user-friendly equation format (similar to that of linear regression) without any pruning or simplification of the model. At the same time, MTFANN maintains similar generalization ability to mono-transfer function networks in a global optimization context. The performance and knowledge extraction of MTFANN were evaluated on a realistic simulation of the Puma 560 robot arm and compared to sigmoid, hyperbolic tangent, linear and sinusoidal networks.

Fault detection in a cold forging process through feature extraction with a neural network

This paper investigates the application of neural networks to the recognition of lubrication defects typical to an industrial cold forging process employed by fastener manufacturers. The accurate recognition of lubrication errors, such as coating not being applied properly or damaged during material handling, is very important to the quality of the final product in fastener manufacture. Lubrication errors lead to increased forging loads and premature tool failure, as well as to increased defect sorting and the re-processing of the coated rod. The lubrication coating provides a barrier between the work material and the die during the drawing operation; moreover it needs be sufficiently robust to remain on the wire during the transfer to the cold forging operation. In the cold forging operation the wire undergoes multi-stage deformation without the application of any additional lubrication. Four types of lubrication errors, typical to production of fasteners, were introduced to a set of sample rods, which were subsequently drawn under laboratory conditions. The drawing force was measured, from which a limited set of features was extracted. The neural network based model learned from these features is able to recognize all types of lubrication errors to a high accuracy. The overall accuracy of the neural network model is around 98% with almost uniform distribution of errors between all four errors and the normal condition.

Improving an inverse model of sheet metal forming by neural network based regression

An inverse model for a sheet meta l forming process aims to determine the initial parameter levels required to form the final formed shape. This is a difficult problem that is usually approached by traditional methods such as finite element analysis. Formulating the problem as a classification problem makes it possible to use well established classification algorithms, such as decision trees. Classification is, however, generally based on a winner-takes-all approach when associating the output value with the corresponding class. On the other hand, when formulating the problem as a regression task, all the output values are combined to produce the corresponding class value. For a multi-class problem, this may result in very different associations compared with classification between the output of the model and the corresponding class. Such formulation makes it possible to use well known regression algorithms, such as neural networks. In this paper, we develop a neural network based inverse model of a sheet forming process, and compare its performance with that of a linear model. Both models are used in two modes, classification mode and a function estimation mode, to investigate the advantage of re-formulating the problem as a function estimation. This results in large improvements in the recognition rate of set-up parameters of a sheet metal forming process for both models, with a neural network model achieving much more accurate parameter recognition than a linear model.

Knowledge extraction from a mixed transfer function artificial neural network

One of the big problems with Artificial Neural Networks (ANN) is that their results are not intuitively clear. For example, if we use the traditional neurons, with a sigmoid activation function, we can approximate any function, including linear functions, but the coefficients (weights) in this approximation will be rather meaningless. To resolve this problem, this paper presents a novel kind of ANN with different transfer functions mixed together. The aim of such a network is to i) obtain a better generalization than current networks ii) to obtain knowledge from the networks without a sophisticated knowledge extraction algorithm iii) to increase the understanding and acceptance of ANNs. Transfer Complexity Ratio is defined to make a sense of the weights associated with the network. The paper begins with a review of the knowledge extraction from ANNs and then presents a Mixed Transfer Function Artificial Neural Network (MTFANN). A MTFANN contains different transfer functions mixed together rather than mono-transfer functions. This mixed presence has helped to obtain high level knowledge and similar generalization comparatively to monotransfer function nets in a global optimization context.

Eccentricity and hardness control in cold rolling mills with a dynamically constructed neural controller

The main objective of a steel strip rolling process is to produce high quality steel at a desired thickness.  Thickness reduction is the result of the speed difference between the incoming and the outgoing steel strip and the application of the large normal forces via the backup and the work rolls.  Gauge control of a cold rolled steel strip is achieved using the gaugemeter principle that works adequately for the input gauge changes and the strip hardness changes.  However, the compensation of some factors is problematic, for example, eccentricity of the backup rolls.  This cyclic eccentricity effect causes a gauge deviation, but more importantly, a signal is passed to the gap position control so to increase the eccentricity deviation.  Consequently, the required high product tolerances are severely limited by the presence of the roll eccentricity effects.
In this paper a direct model reference adaptive control (MRAC) scheme with dynamically constructed neural controller was used.  The aim here is to find the simplest controller structure capable of achieving an optimal performance.  The stability of the adaptive neural control scheme (i.e. the requirement of persistency of excitation and bounded learning rates) is addressed by using as the inputs to the reference model the plant's state variables.  In such a case, excitation is due to actual plant signals (states) affected by plant disturbances and noise.  In addition, a reference model in the form of a filter with a desired transfer function using Modulus Optimum design was used to ensure variance in the desired dynamic characteristics of the system.  The gradually decreasing learning rate employed by the neural controller in this paper is aimed at eliminating controller instability resulting from over-aggressive control.  The moving target problem (i.e. the difficulty of global neural networks to perfrom several separate computational tasks in closed -loop control) is addressed by the localized architecture of the controller.  The above control scheme and learning algorithm offers a method for automatic discovery of an efficient controller.
The resulting neural controller produces an excellent disturbance rejection in both cases of eccentricity and hardness disturbances, reducing the gauge deviation due to eccentricity disturbance from 33.36% to 4.57% on average, and the gauge deviation due to hardness disturbance from 12.59% to 2.08%.