Neuroscience: Pattern Recognition with Neural Networks

Neuroscience: Pattern Recognition with Neural Networks

Retinal ganglion cells, that is, the output cells of the retina, have concentric on-center or off-center receptive fields. The on-center receptive field consists of an excitatory circular region surrounded by an inhibitory annular region, while the off-center receptive field has the opposite structure. These cells are supposed to extract contrast components from patterns projected onto the retina. Photoreceptors and ganglion cells are not directly connected in an actual biological retina. There are several kinds of cells in between and the information is processed by the interactions of such cells. Engineers in pattern recognition often classify the process of self-organization into supervised learning (or learning-with-a-teacher) and unsupervised learning (or learning-without-a-teacher). A famous example of a classical network that can be trained by supervised learning is the three-layered perceptron. Although the perceptron has a three-layered hierarchical structure, it is only in the highest stage that the cells have variable input connections.

Read More

Psychology, Neurobiology and Modeling: The Science of Hebbian Reverberations

Psychology, Neurobiology and Modeling: The Science of Hebbian Reverberations

This chapter discusses introspective considerations for the necessity of having attractors, which neuropsychology has termed as reverberations. There exist experimental evidences for reverberations and attractors exist and are observable in neurophysiology. Attractors that are observed neurophysiologically, which appear to be required for mental processing, are very informative both for describing the functioning of the brain and for measuring information about computation processes. There are other phenomena in cognitive psychology whose interpretation in terms of correlated attractors is quite simple. For example, there is a well known phenomenon called false alarm in which subjects are shown a set of objects and then a sublist is defined. Patterns are shown very quickly, at random, and the subject has to respond after each presentation whether the stimulus belongs to the sublist. If the pattern shown is correlated with one belonging to the sublist, then the number of errors grows. In attractor language, it happens because each one of the patterns carries information of the fact that it was seen very often in temporal proximity to a pattern that belongs to the sublist.

Read More

Neuroscience: Towards a Kinetic Theory of Cortical-like Neural Fields

Neuroscience: Towards a Kinetic Theory of Cortical-like Neural Fields

The kinetic theory of neural fields is based fundamentally on the description of electrical properties of single neurons and on some statistical hypotheses about their connecting network. This chapter discusses the basis of this theory and its main results. It discusses a basilar, minimal set of biological neural features, except those linked to neural plasticity. Such characteristics are utilized for the construction, through statistical-physics methods, of a general mathematical model, which does not refer to particular brain regions. The chapter discusses some results of the application of the theory to cortical-like neural fields and to attention and learning. Neural systems are considered as finite regions of the three-dimensional space filled with two different types of particles. Neurons, characterized by an inner excitation corresponding to the subthreshold membrane potential, constitute a kind of motionless, solid background. They, under some conditions, emit impulses in space.

Read More

Neuroscience: Qualitative Overview of Population Neurodynamics

Neuroscience: Qualitative Overview of Population Neurodynamics

The estimates of the number of neurons in a cubic millimeter of cortex range from 104 to 106 and from 103 to 105 for the number of synaptic connections on each dendritic tree. The typical sensory input for a conditioned stimulus is carried by a parallel array of an immeasurably large number of receptor axons and the motor outflow for a typical conditioned response, a fraction of a second later, is carried in parallel by an immeasurably large number of motor axons. The networks of model neurons are constructed by induction that represent the actions of a small subset of neurons on the implicit assumption that what the other (unobserved) neurons are doing during a perceptual act or a conditioned reflex is not important for the activity of the observed subset. This approach might be called statistical mechanics of brain function, comparable to the study of molecules in an ideal gas in terms of their kinetic energy, position, and collision rates.

Read More

Neuroscience:Noise and Chaos in Neural Systems

Neuroscience:Noise and Chaos in Neural Systems

This chapter discusses two questions (1) how to characterize the dendritic arborization patterns and what could be the mechanism of the generation of dendritic branching trees and (2) what can be the role of environmental noise in the ontogenetic formation of ordered cortical structures. The dendritic arborization can be classified based on (1) randomness versus regularity in regard to branching of the various dendritic segments, resulting in a continuous spectrum ranging from the “radiate” to the “tufted” types of arborization and (2) the degree of deviation of the individual branches from a radiated and rectilinear course. A three-variable extension of the Turing-Gierer-Meinhardt (TGM) model describes network formation based on the diffusible morphogen concept. This model leads to uniform behavior. Accepting the view that neural dendritic branching might have a self-similar character, its development can be modeled by fractal generating growth models.

Read More

Neuroscience: Diffusion Models of Single Neurones' Activity

Neuroscience: Diffusion Models of Single Neurones' Activity

This chapter presents a survey of one-dimensional stochastic diffusion models for the membrane potential of a single neuron, with emphasis on the probabilistic properties of the models and on the related first-passage-time (FPT) problems, namely, on the determination of the neuronal output. It discusses how the much celebrated Ornstein-Uhlenbeck (OU) neuronal model can be obtained as the limit of a Markov process with discrete state space in continuous time. The stochastic models of neuronal activity rank among the most advanced applications of the theory of stochastic processes to biology. There are several reasons why many of these neuronal models stem out of the theory of diffusion processes: (1) the theory of diffusion processes is well developed and hence, it allows the neural modelers to apply many general results to the specific application area, (2) for neurons with many synaptic inputs, there is a rather good correspondence between the models and the biological reality, (3) even when some of the assumptions underlying the model may be somewhat questionable, the output behavior of the neuron can often be well approximated by that of a suitably chosen diffusion process, which is, for instance, strikely true for the well-known Gerstein-Mandelbrot diffusion model, and (4) from a biophysical point of view, the neuronal models of single cells reflect the electrical properties of the membrane via electric circuit models that contain energy storage elements.

Read More

Neuroscience: Mechanisms Responsible for Epilepsy in Hippocampal Slices Predispose the Brain to Collective Oscillations

Neuroscience: Mechanisms Responsible for Epilepsy in Hippocampal Slices Predispose the Brain to Collective Oscillations

About 90% of the neurons in the CA3 region are pyramidal cells, the principal neurons that supply CA3 output to other parts of the brain (including, in particular, the CA1 hippocampal region and the opposite CA3 region—there is one hippocampus in each cerebral hemisphere). Of particular relevance to epilepsy and probably to normal brain function is the fact that CA3 neurons synaptically excite one another. These connections are called recurrent. The remaining neurons, generally nonpyramidal in shape, are probably all inhibitory and most or all of them use γ-aminobutyric acid (GABA) as their transmitter. Pyramidal neurons have two sets of dendrites, called basilar and apical, with recurrent synaptic connections into both sets of dendrites. Presynaptic neurons release transmitters onto a given postsynaptic neuron causing membrane synaptic currents to flow and possibly initiating intracellular secondary metabolic processes that in turn influence the membrane. There are also synaptic receptors whose primary function is to couple to intracellular biochemical processes. As metabolic processes are, in general, slower than voltage-gated membrane processes, the different receptor coupling mechanisms introduce a range of time scales into neuronal function.

Read More