Neural Process Models Research Articles

Implementing Spiking Neural Networks for Real-Time Signal-Processing and Control Applications: A Model-Validated FPGA Approach

In this paper, we present two versions of a hardware processing architecture for modeling large networks of leaky-integrate-and-fire (LIF) neurons; the second version provides performance enhancing features relative to the first. Both versions of the architecture use fixed-point arithmetic and have been implemented using a single field-programmable gate array (FPGA). They have successfully simulated networks of over 1000 neurons configured using biologically plausible models of mammalian neural systems. The neuroprocessor has been designed to be employed primarily for use on mobile robotic vehicles, allowing bio-inspired neural processing models to be integrated directly into real-world control environments. When a neuroprocessor has been designed to act as part of the closed-loop system of a feedback controller, it is imperative to maintain strict real-time performance at all times, in order to maintain integrity of the control system. This resulted in the reevaluation of some of the architectural features of existing hardware for biologically plausible neural networks (NNs). In addition, we describe a development system for rapidly porting an underlying model (based on floating-point arithmetic) to the fixed-point representation of the FPGA-based neuroprocessor, thereby allowing validation of the hardware architecture. The developmental system environment facilitates the cooperation of computational neuroscientists and engineers working on embodied (robotic) systems with neural controllers, as demonstrated by our own experience on the Whiskerbot project, in which we developed models of the rodent whisker sensory system.

Glial cells for information routing?

We investigate a possible functional role of glial cells as information routing devices of the cerebral cortex. On the one hand, functionally motivated models of neural information processing were lately suggested which rely on short-term changes of connections between neural modules to dynamically route neural activity. Although successful in practice, the routing mechanisms of these models require synaptic efficacy control of large sets of synapses that is difficult to implement neurally. On the other hand, recent experiments show an active role of glial cells (astrocytes) in the interaction with large numbers of synapses. Astrocytes are sensitive to neurotransmitters released by the presynaptic terminal and in turn can influence synaptic efficacy by release of so called gliotransmitters. An analysis of the most recent literature shows that glial cells are a well-suited and natural candidate for the implementation of information routing mechanisms.

Inertia and memory in ambiguous visual perception

Perceptual multistability during ambiguous visual perception is an important clue to neural dynamics. We examined perceptual switching during ambiguous depth perception using a Necker cube stimulus, and also during binocular rivalry. Analysis of perceptual switching time series using variance-sample size analysis, spectral analysis and time series shuffling shows that switching times behave as a 1/f noise and possess very long range correlations. The long memory feature contrasts sharply with the traditional satiation models of multistability, where the memory is not incorporated, as well as with recently published models of multistability and neural processing, where memory is excluded. On the other hand, the long memory feature favors the concept of "dynamic core" or coalition of neurons, where neurons form transient coalitions. Perceptual switching then corresponds to replacement of one coalition of neurons by another. The inertia and memory measures the stability of a coalition: a strong and stable coalition has to be won over by another similarly strong and stable coalition, resulting in long switching times. The complicated transient dynamics of competing coalitions of neurons may be addressable using a combination of functional imaging, measurement of frequency-tagged magnetoencephalography and frequency-tagged encephalography, simultaneous recordings of groups of neurons in many areas of the brain, and concepts from statistical mechanics and nonlinear dynamics theory.

Trying to See Red Through Stickleback Photoreceptors: Functional Substitution of Receptor Sensitivities

AbstractA key to understanding animal behavior is knowledge of the sensory information animals extract from their environment. For visually motivated tasks, the information animals obtain through their eyes is often assumed to be essentially the same as that perceived by humans. However, known differences in structure and processing among the visual systems of different animals clearly indicate that the world seen by each is different. A well‐characterized difference between human and other animal visual systems is the number of types and spectral sensitivities of their photoreceptors. We are developing a technique, functional substitution, that exploits knowledge of these differences to portray for human subjects, colors as they would appear through the photoreceptors of another animal. In a specific application, we ask human subjects to rank hues of male threespine stickleback (Gasterosteus aculeatus) throats viewed through stickleback photopigments. We compare these ranks to ranks of the same throat hues viewed through normal human photoreceptors. We find essentially no difference between the two sets of rankings. This suggests that any differences in human and stickleback rankings of such hues would result from differences in post‐receptoral neural processing. Using a previously developed model of stickleback neural processing, we established another ranking of the hues which was again essentially the same as the rankings produced by the human subjects. A growing literature indicates that stickleback do rank such hues in the evaluation of males as potential mates or threats. Although our results do not demonstrate that humans and stickleback use the same mechanisms to assess color, our experiments significantly failed to show that stickleback and human rankings of throat hues should be different. Nevertheless, a comparison of all these rankings to ranks derived from subjective color scoring by human observers suggests that color scoring may utilize other cues and should thus be used cautiously.

On levels of cognitive modeling

The article first addresses the importance of cognitive modeling, in terms of its value to cognitive science (as well as other social and behavioral sciences). In particular, it emphasizes the use of cognitive architectures in this undertaking. Based on this approach, the article addresses, in detail, the idea of a multi-level approach that ranges from social to neural levels. In physical sciences, a rigorous set of theories is a hierarchy of descriptions/explanations, in which causal relationships among entities at a high level can be reduced to causal relationships among simpler entities at a more detailed level. We argue that a similar hierarchy makes possible an equally productive approach toward cognitive modeling. The levels of models that we conceive in relation to cognition include, at the highest level, sociological/anthropological models of collective human behavior, behavioral models of individual performance, cognitive models involving detailed mechanisms, representations, and processes, as well as biological/physiological models of neural circuits, brain regions, and other detailed biological processes.

"Seeing" in the dark: neuromorphic VLSI modeling of bat echolocation

While modern sonar and radar processing systems are constructed using large sensor arrays that have incredible imaging and localization capabilities, the question of what bats experience and how they process echo data with only two ears and a pea-sized brain remains a major mystery. The exploration of this question and the attempt to construct a functional model of the bat's neural signal processing using neuromorphic very large scale integration (VLSI) techniques and robotics have provided an interesting framework for our laboratory's research program.

A face/object recognition system using coarse region segmentation and dynamic-link matching

An image recognition model that combines some neural network-based image processing models is proposed. The recognition procedure consists of coarse region segmentation/extraction performed by a resistive-fuse network, Gabor wavelet transformation (GWT) and dynamic-link matching. We have also developed a PC-based face/object recognition system including FPGA implementation of the resistive-fuse network. The system has successfully achieved real-time face recognition from a natural scene image.

A diffusion-neural-network for learning from small samples

Neural information processing models largely assume that the patterns for training a neural network are sufficient. Otherwise, there must exist a non-negligible error between the real function and the estimated function from a trained network. To reduce the error, in this paper, we suggest a diffusion-neural-network (DNN) to learn from a small sample consisting of only a few patterns. A DNN with more nodes in the input and layers is trained by using the deriving patterns instead of original patterns. In this paper, we give an example to show how to construct a DNN for recognizing a non-linear function. In our case, the DNN’s error is less than the error of the conventional BP network, about 48%. To substantiate the special case arguments, we also study other two non-linear functions with simulation technology. The results show that the DNN model is very effective in the case where the target function has a strong non-linearity or a given sample is very small.

Fault detection and classification in chemical processes based on neural networks with feature extraction

Abstract Feed forward neural networks are investigated here for fault diagnosis in chemical processes, especially batch processes. The use of the neural model prediction error as the residual for fault diagnosis of sensor and component is analyzed. To reduce the training time required for the neural process model, an input feature extraction process for the neural model is implemented. An additional radial basis function neural classifier is developed to isolate faults from the residual generated, and results are presented to demonstrate the satisfactory detection and isolation of faults using this approach.

From Neurophysiology to Neural Nets

The study of the function of excitable tissues in response to electrical signals constitutes the basis for understanding and reproducing phenomena of specific neural systems. Of course, the collection of vast numbers of experimental data, enriched by experience, helped to derive accurate mathematical models of neural processing, which thereby contributed to the design of artificial models of biological neural networks. The most important milestone regarding this field of study was the seminal work performed by the Nobel prize winners Hodgkin and Huxley in 1952, who, by their work, captured the characteristics of the active behaviours of neural membranes and their threshold dependent nonlinear dynamics. Since then, several outstanding mathematical models of neural signals and their propagation in active and nonlinear mediums have been developed. The term “artificial neural nets” is currently used to mean all the architecture, from software to hardware that is built on biological understandings and devoted to the tasks typically performed by living creatures. These tasks include learning from examples, generalisation ability, pattern recognition, memory and so on. Another increasingly important field is the specific design of biologically inspired neural architectures that are able to solve specific important tasks such as locomotion generation and control. We report here on our recent research activity in this direction.

A neural network model-based open-loop optimization for the automated thermoplastic composite tow-placement system

This study demonstrates the use of on-line optimization algorithms to calculate optimum process set points for manufacturing processes such as the automated thermoplastic tow-placement (ATP) system. An in situ non-linear optimization technique based on artificial neural networks has been developed. This method is implemented in the ATP process and utilizes neural network-based process models to predict material quality as a function of process set points. The set points are computed by maximizing the throughput and maintaining a desired minimum quality. Process history can greatly affect the final part quality and, therefore, is an integral part of the optimization. The controller is validated for the highly non-linear ATP process and successfully predicts optimum processing parameters. The developed approach is applicable to many other manufacturing processes where process simulations exist and conventional control techniques are lacking.

A computer model of neural processes observed in the cat motor cortex during performance of an operant movement.

This report describes a computer model of a "column" in the cat motor cortex. The model includes two layers of two-segment pyramidal neurons with two groups of inhibitory interneurons in each layer, which selectively control the somatic and dendritic segments of the pyramidal cells. In this model, neurons include active sodium, calcium, and several types of potassium currents. Excitatory connections between neurons are of the AMPA and NMDA types, while collateral connections between neurons of the upper layer are mainly of the NMDA type; connections between neurons in the lower layer are of the AMPA type. All inhibitory connections are of the GABA(A) type. The model reproduces the main neuronal processes seen in the cat motor cortex during performance of an operant movement. Pyramidal neurons of the upper layer generate primary and secondary responses to external stimuli. As in real experiments, secondary NMDA-dependent responses appear when GABA(A) inhibition is weakened and disappear when stimulation is increased; these properties of secondary responses are only reproduced when NMDA receptors are located in the terminals of collateral connections. Using only rapid NMDA-independent connections, neurons in the lower layer generate a slow bell-shaped wave of excitation (a "motor command"), which is formed by sequential activation of neurons with dendritic trees of different sizes.

Trajectory tracking of a batch polymerization reactor based on input–output-linearization of a neural process model

Input–output-linearization via state feedback offers the potential to serve as a practical and systematic design methodology for nonlinear control systems. Nevertheless, its widespread use is delayed due to the fact that developing an accurate plant model based on physical principles is often too costly and time consuming. Data-based modeling of dynamic systems using neural networks offers a cost-effective alternative. This work describes the methodology of input–output-linearization using neural process models and gives an extended simulative case study of its application to trajectory tracking of a batch polymerization reactor.

Princeton Physicist Garners Dirac Medal

John J. Hopfield, a professor of computational neurobiology and biophysics at Princeton University, was awarded the Dirac Medal on 8 August by the Abdus Salam International Centre for Theoretical Physics in Trieste, Italy. With the award, the ICTP has recognized, for the first time, the very different mode of theory that is necessary for making the connection between physics and biology.Each year on 8 August, ICTP commemorates the birthday of physicist and Nobel Prize winner P. A. M. Dirac by presenting the medal in recognition of significant contributions to theoretical physics and mathematics. This year’s medal acknowledges Hopfield’s “important contributions in an impressively broad spectrum of scientific subjects.” The ICTP points out that his “special and rare gift is his ability to cross the interdisciplinary boundary to discover new questions and propose answers that uncover the conceptual structure behind the experimental facts.”Hopfield’s achievements include the Hopfield model of neural processing, which demonstrates how computation is qualitatively different in a computer compared with that in the brain and relates some of the brain’s “neural algorithms” to the mathematics of magnetic systems. More recently, according to the ICTP, “he has found an entirely different [collective] organizing principle in olfaction and demonstrated a new principle in which neural function can take advantage of the temporal structure of the ‘spiking’ interneural communication.”In addition to the medal, Hopfield received a prize of $5000.© 2001 American Institute of Physics.

Afferents from the auditory thalamus synapse on inhibitory interneurons in the lateral nucleus of the amygdala

Physiological studies suggest that afferents to the lateral nucleus of the amygdala (LA) from the auditory thalamus initiate feedforward inhibition [Li et al. (1996b)]. This model of neural processing requires that thalamic afferents synapse directly onto inhibitory interneurons. To determine whether such synaptic contacts occur, we combined anterograde tract tracing with interneuron immunocytochemistry. The anterograde tracer biotinylated dextran amine (BDA) was injected into the auditory thalamus. Inhibitory interneurons in the LA were identified using antibodies directed against gamma aminobutyric acid (GABA) or one of the calcium binding proteins (CBPs), parvalbumin (PARV), calbindin (CALB), or calretinin (CALR), since CBPs identify distinct populations of GABAergic cells within the amygdala. The distribution of GABAergic and CBP interneurons in each subregion of the LA was examined by light microscopy and the relationships between thalamo-amygdala terminals and interneurons were examined by confocal and electron microscopy. Immunoreactive cells were distributed in all three subdivisions of LA, except for CALR-ir neurons, which were sparse in the dorsal subregion and were found mainly in the ventromedial and ventrolateral subregions. Confocal microscopy revealed some thalamo-amygdala terminals in close proximity to LA interneurons, while electron microscopy showed that thalamo-amygdala terminals made direct synaptic contacts onto distal dendritic processes of inhibitory neurons. These data provide morphological evidence that thalamic afferents synapse directly onto inhibitory interneurons in LA, and are consistent with the possibility that inputs from the auditory thalamus initiate feedforward inhibition in LA. This architecture could play an important role in the suppression of background neural noise, thereby enhancing the response of LA cells to incoming auditory stimuli. Synapse 38:124–137, 2000. © 2000 Wiley-Liss, Inc.

A neural model of medial geniculate body and auditory cortex detecting target distance independently of target velocity in echolocation

Using frequency-modulated (FM) ultrasonic pulses, echolocating bats detect with a high accuracy the arrival time differences between emitted pulses and their echos generated by targets, even though the targets are moving. We propose a neural mechanism by which the bat may detect the time differences independently of the target velocity using FM ultrasonic pulses. The neural processing model presented consists of the networks in MGB for calculating the temporal correlation between pulse and echo for each frequency and the network of AC for integrating the calculated results over the frequency range used by bat. We found that the key factor generating the detection independent of target velocity is the value of frequency modulation rate of about 6 kHz.

The patchwork engine: image segmentation from shape symmetries

We propose blind segmentation of images into shape-related ‘patches’ based on pre-calculated local symmetries (Van Tonder, G.J. & Ejima, Y. (1999). (Forthcoming a) Flexible computation of shape symmetries. Submitted for publication) in shape boundary contours. First, lateral weights between all points in the boundary contour map are assigned analogous to Euclidean distance maps in watershed segmentation (Beucher, S. & Lantejoul, C. (1979). Use of watersheds in contour detection. Proceedings of the International Workshop on Image Processing, CCETT, Rennes, France.). Lateral weights are then used to: (1) extract local maxima in symmetries; (2) link maxima within locally enclosed boundary contours; and (3) reconstruct shape contours using symmetry maxima as ‘seeds’. The new model overcomes weaknesses of watershed segmentation. The new model closes gaps in relatively more solid image contours, but it is fundamentally different from methods based on contour interpolation (Grossberg, S., Mingolla, E. & Todorovć, D. (1989). A neural network architecture for pre-attentive vision, IEEE Transactions on Biomedical Engineering 36, 65–84; Heitger, F. & von der Heydt, R. (1993). A computational model of neural contour processing: figure-ground segregation and illusory contours. Proceedings of the Fourth International Conference on Computer Vision, IEEE Computer Society Press, Washington D.C. (pp. 32–40)). Images are segmented into shape-relevant color-by-number-like patches which compare well to related methods (Gauch, J. & Pizer, M. (1993). The intensity axis of symmetry and its application to image segmentation, IEEE Transactions on Pattern Analysis and Machine Intelligence, 15 (8), 753–770; Ilg, W. & Ogniewicz, R. (1995). The application of Voronoi skeletons to perceptual grouping in line images, Proceedings of the 11th International Conference on Pattern Recognition, The Hague, The Netherlands, pp. 382–385; Zhu, S.C. & Yuille, A.L. (1996) FORMS: a flexible object recognition and modeling system, International Journal of Computer Vision, 20 (3), 187–212.). Two primitive operations, comparison and merging of patches, are proposed as drives for exposing more global shape contours from patches. We conclude that symmetry goes beyond abstract shape morphology: it can contribute to figure-ground segmentation in early vision and form part of primitive operations needed to create hypotheses of complex shape.

Four-Dimensional Neuronal Signaling by Nitric Oxide: A Computational Analysis

Nitric oxide (NO) is now recognized as a transmitter of neurons that express the neuronal isoform of the enzyme nitric oxide synthase. NO, however, violates some of the key tenets of chemical transmission, which is classically regarded as occurring at points of close apposition between neurons. It is the ability of NO to diffuse isotropically in aqueous and lipid environments that has suggested a radically different form of signaling in which the transmitter acts four-dimensionally in space and time, affecting volumes of the brain containing many neurons and synapses. Although "volume signaling" clearly challenges simple connectionist models of neural processing, crucial to its understanding are the spatial and temporal dynamics of the spread of NO within the brain. Existing models of NO diffusion, however, have serious shortcomings because they represent solutions for "point-sources," which have no physical dimensions. Methods for overcoming these difficulties are presented here, and results are described that show how NO spreads from realistic neural architectures with both simple symmetrical and irregular shapes. By highlighting the important influence of the geometry of NO sources, our results provide insights into the four-dimensional spread of a diffusing messenger. We show for example that reservoirs of NO that accumulate in volumes of the nervous system where NO is not synthesized contribute significantly to the temporal and spatial dynamics of NO spread.

Afferents from the auditory thalamus synapse on inhibitory interneurons in the lateral nucleus of the amygdala.

Physiological studies suggest that afferents to the lateral nucleus of the amygdala (LA) from the auditory thalamus initiate feedforward inhibition [Li et al. (1996b)]. This model of neural processing requires that thalamic afferents synapse directly onto inhibitory interneurons. To determine whether such synaptic contacts occur, we combined anterograde tract tracing with interneuron immunocytochemistry. The anterograde tracer biotinylated dextran amine (BDA) was injected into the auditory thalamus. Inhibitory interneurons in the LA were identified using antibodies directed against gamma aminobutyric acid (GABA) or one of the calcium binding proteins (CBPs), parvalbumin (PARV), calbindin (CALB), or calretinin (CALR), since CBPs identify distinct populations of GABAergic cells within the amygdala. The distribution of GABAergic and CBP interneurons in each subregion of the LA was examined by light microscopy and the relationships between thalamo-amygdala terminals and interneurons were examined by confocal and electron microscopy. Immunoreactive cells were distributed in all three subdivisions of LA, except for CALR-ir neurons, which were sparse in the dorsal subregion and were found mainly in the ventromedial and ventrolateral subregions. Confocal microscopy revealed some thalamo-amygdala terminals in close proximity to LA interneurons, while electron microscopy showed that thalamo-amygdala terminals made direct synaptic contacts onto distal dendritic processes of inhibitory neurons. These data provide morphological evidence that thalamic afferents synapse directly onto inhibitory interneurons in LA, and are consistent with the possibility that inputs from the auditory thalamus initiate feedforward inhibition in LA. This architecture could play an important role in the suppression of background neural noise, thereby enhancing the response of LA cells to incoming auditory stimuli.

Using neural networks to construct models of the molecular beam epitaxy process

This paper presents the systematic characterization of the molecular beam epitaxy (MBE) process to quantitatively model the effects of process conditions on film qualities. A five-layer, undoped AlGaAs and InGaAs single quantum well structure grown on a GaAs substrate is designed and fabricated. Six input factors (time and temperature for oxide removal, substrate temperatures for AlGaAs and InGaAs layer growth, beam equivalent pressure of the As source and quantum well interrupt time) are examined by means of a fractional factorial experiment. Defect density, X-ray diffraction, and photoluminescence are characterized by a static response model developed by training back-propagation neural networks. In addition, two novel approaches for characterized reflection high-energy electron diffraction (RHEED) signals used in the real-time monitoring of MBE are developed. In the first technique, principal component analysis is used to reduce the dimensionality of the RHEED data set, and the reduced RHEED data set is used to train neural nets to model the process responses. A second technique uses neural nets to model RHEED intensity signals as time series, and matches specific RHEED patterns to ambient process conditions. In each case, the neural process models exhibit good agreement with experimental results.