Search

The topic of artificial consciousness asks the following question. Were a manufactured object appear to be conscious what mechanisms could be responsible for such a state of affairs? First the paper defines the foundation of the theory as a postulate which requires the manufactured object to be neural, specifying that the firing of someneurons is responsible for conscious states. This leads to the title ‘neuroconsciousness’. It follows this up by a series of 12 corollaries which is open-ended in the belief that a definition of consciousness is an open-ended series of properties. The method of theoretical expression is that of a neural state machine for which the state structure is the product of learning and generalisation. Central to this scheme is a novel concept of ‘iconic’ learning which creates states that are sampled versions of sensory experience. The paper ends with both a technical critique that is directed towards the construction of this type of machinery and a philosophical critique which asks how helpful is the concept of artificial consciousness in elucidating the real thing.

The convergence-zone model shows how sparse, random memory patterns can lead to one-shot storage and high capacity in the hippocampal component of the episodic memory system. This paper presents a biologically more realistic version of the model, with continuously weighted connections and storage through Hebbian learning and normalization. In contrast to the gradual weight adaptation in many neural network models, episodic memory turns out to require high learning rates. Normalization allows earlier patterns to be overwritten, introducing time-dependent forgetting similar to the hippocampus.

Recent evidence suggests that (a) auditory cortical neurons are tuned to complex time-varying acoustic features, (b) auditory cortex consists of several fields that decompose sounds in parallel, (c) the metric for such decomposition varies across species, and (d) auditory cortical representations can be rapidly modulated. Past computational models of auditory cortical processing cannot capture such representational complexity. This paper proposes a novel framework in which auditory signal processing is characterized as an adaptive transformation from a one-dimensional space into an n-dimensional auditory parameter space. This transformation can be modeled as a chirplet transform implemented via a self-organizing neural network.

Capacity limited memory systems need to gradually forget old information in order to avoid catastrophic forgetting where all stored information is lost. This can be achieved by allowing new information to overwrite old, as in the so-called palimpsest memory. This paper describes a new such learning rule employed in an attractor neural network. The network does not exhibit catastrophic forgetting, has a capacity dependent on the learning time constant and exhibits recency effects in retrieval.

This paper studies dynamical properties of spatio-temporal pattern sequences (‘synfire chains’) in associative networks of spiking neurons. Employing postsynaptic potentials with a finite rise time, the replay speed of stored sequences can be controled by unspecific background signals. In addition, the speed also depends on the number of co-activated sequences, but balanced inhibition can prevent this dependency. An implicit equation is derived and solved numerically which relates the speed to the network parameters. Simulations reveal instabilities for low and high control signals, which are traced back to four different destabilizing mechanisms.

This paper presents simulations of a small patch of the turtle's visual cortex with 20 pyramidal cells and four smooth stellate cells. Our model captures the basic geometry and temporal structure of the visual cortex and has been shown to generate propagating waves of activity that contain information about a moving stimulus.

This paper describes a mathematical model of the axon of the cerebellar granule cell as it ascends into the molecular layer and branches into its two parallel fiber segments. We present a theoretical and computational analysis of the input resistance at the branching point and the degree of signal attenuation between the ascending segment of the axon and the parallel fibers. The results suggest considerable passive signal decay in this branch point. If correct, this result has important implications for cerebellar function, as granule cell activation would be expected to always propagate in the ascending but not necessarily the parallel fiber segments of the same axon.

While several methods have been developed to estimate the Hurst exponent Hfor self-affine signals, they are not sufficiently accurate when the signal has a large harmonic component. Similarly, the frequency of harmonic components cannot be estimated accurately from signals which have an underlying power-law spectrum. Coarse graining spectral analysis (CGSA) has been suggested to be capable of separating random fractals from harmonic components. In this paper, we assess the robustness and effectiveness of the CGSA method and find that it is best suited for separating harmonic and fractal components from mixed time series with low-frequency harmonics and long-term positive correlations.

With a model of the prefrontal cortical circuit, proposed here, this paper investigates the post-cue (delay-period) activity of the prefrontal cortical neurons for a spatial working memory task. The computer simulations of the model circuit show that a slight change in the strength of the local inhibition dramatically alters the post-cue activity as well as the tuning properties of the model prefrontal cortical neurons. Appropriate regulation of the local inhibition is necessary to form and maintain spatial working memory. These results suggest the capability of the local inhibition to play an important role in the active control of working memory.

Recently, it has been shown that cerebellar Purkinje cell can acquire (memorize) the signals, which come to the climbing fiber cell from external sources [4,5]. The memorizing mechanism is based on using phase of climbing fiber firing. In preceding papers two major points were unclear: (1)how the system behaves in a case of non-stationary signals, and (2)how information transformation in cells of cerebellar nuclei affects model behavior. Herewith in computer simulation experiments, we demonstrate that the phase-based cerebellar system is well suited for acquisition of non-stationary signals. The system is also capable to perform signal interpolation as well as acquisition of randomly presented relations. Mathematical analysis, supported with computer simulations, also demonstrate that signal transformation in cerebellar nuclei is consistent with the proposed model and yields same results as the simplified network.

Computational models and intracellular recordings were combined to characterize synaptic bombardment in neocortical pyramidal neurons during active states in vivo. Application of tetrodotoxin revealed that synaptic activity accounts for up to 80% of the input conductance. These experiments were replicated assuming that excitatory and inhibitory synapses release randomly at high rates (1–5Hz). A significant correlation between synaptic events had to be introduced to account for the membrane potential fluctuations recorded experimentally. We conclude that pyramidal neurons in vivo experience high-frequency synaptic events with a significant correlation. The consequences of this synaptic bombardment on cellular responsiveness are investigated in the companion paper.

This paper analyzes the distribution of both local and callosal connections in cat visual cortex V1. Using a computational model it is shown that the pattern of intra-hemispherical connections can be accounted for by assuming that the connection probability between two cells is dependent primarily on two factors — receptive field overlap and orientation selectivity. These factors can also account for many experimental observations of callosal connectivity patterns. These include the extent of the ipsilateral representation of the visual field in the ipsilateral hemisphere and the distribution of callosally projecting and receiving cells along the V1/V2 border. The findings indicate that common mechanisms underly the development of connections within and between the two hemispheres.

A convergence theorem for the Cascade Correlation is described and based on this, results on convergence speed of the algorithm are obtained. It is possible to forecast the cost when adding hidden units: this allows us to understand if the Cascade Correlation works properly. A critical test is presented where the growing process must be modified in order to avoid unsatisfactory outcome; a parameter for detecting such occurrences is proposed and therapies are suggested. The results obtained in this paper may be used as a kernel for building new types of growing networks, whose topology may be better matched with the training data.

In this paper, we present 3 different neural network-based methods to perform variable selection. OCD — Optimal Cell Damage — is a pruning method, which evaluates the usefulness of a variable and prunes the least useful ones (it is related to the Optimal Brain Damage method of Le Cun et al.). Regularization theory proposes to constrain estimators by adding a term to the cost function used to train a neural network. In the Bayesian framework, this additional term can be interpreted as the log prior to the weights distribution. We propose to use two priors (a Gaussian and a Gaussian mixture) and show that this regularization approach allows to select efficient subsets of variables. Our methods are compared to conventional statistical selection procedures and are shown to significantly improve on that.

Starting with a biological model of the foveal and parafoveal regions of the retina, this paper shows that simple considerations about the spatiotemporal filtering processed by the retinal neural network can account for colour processing at the level of early vision. We establish that the Red, Green, Blue signal can be considered as a low-pass luminance signal plus a colour-modulated signal. The structure of the Outer- and Inner Plexiform Layers of the retina leads to spatial low- and high-pass filters which account for the achromatic and transient characteristics of the Y ganglion cells as well as for the spatiotemporal colour-opponent properties of X ganglion cells. Considering this property and the logarithmic transduction of photoreceptors, it is easy to postulate that, after the retina at the level of the Lateral Geniculate Nucleus, a simple low-pass filtering can pave the way to the well known colour constancy phenomenon.

This paper presents a novel unsupervised learning paradigm for feature extraction in linear networks under the constraint that no information distortion occurs in the input-output map. Independent feature extraction is then accomplished by minimizing an appropriate information theory based measure of the statistical dependence between the output components. In the case of Gaussian input distribution, a learning rule which preserves the rotation property of the input-output map is derived by using the Lyapunov function type approach. This learning rule is shown to converge to the solution of the standard PCA. Different learning rules are derived for the case of an arbitrary input distribution. They are based on the Edgeworth expansion of a probability density function as well as on the suitably chosen parameterization of admissable linear input/output maps. Examples which validate the introduced learning paradigm are presented.

In this paper we will compare the performance of two techniques (the Maximum Likelihood Estimation (MLE) and the Population Vector (PV)) for estimating the interpretation of neuronal activity in a population of neurons. Although such a comparison has been made before, so far only homogeneous distributions of receptive fields have been investigated. Since the performance of both methods depends on the distribution of the receptive fields we have tested the performance for homogeneous and inhomogeneous distributions. The results demonstrate that in general the MLE method outperforms the Population Vector. However, the MLE method depends heavily on the shape of the receptive field properties of the neurons, which is not the case for the PV method. Moreover, the MLE method may give rise to artefactual results for inhomogeneous distributions of receptive fields. For the PV method the shape of the receptive field is not as important. Moreover, for the Population Vector the optimal width of the receptive field remains more or less constant when the decrease in density is small relative to the optimal width. In this case the information decreases proportionally with the density of receptive fields.