Research in Computer Science and Informatics at the University of Portsmouth

Additional information examples: Plymouth

OtherDetails (Plymouth)
NMDA synaptic currents are assumed to be highly important for the active maintenance of items in working memory by providing a relatively constant recurrent synaptic drive. This paper establishes a new role for NMDA currents, through physiological recordings, nonlinear data analysis, and computational modeling, that reaches far beyond their potential role in maintaining higher levels of spiking activity. Rather NMDA inputs control the dynamical regime and spiking patterns of postsynaptic neurons, and in particular may lift prefrontal cortex neurons into a chaotic irregular regime. These findings should have highly important implications for temporal sequence processing and neural coding.
This paper presents a novel biologically-inspired, hardware-realisable spiking neuron model, called the Temporal Noisy-Leaky Integrator (TNLI). The dynamic applications of the model as well as its applications in Computational Neuroscience are demonstrated and a learning algorithm based on postsynaptic delays is proposed. Overall this work illustrates how a hardware-realisable neuron model can capitalise on the unique computational capabilities of biological neurons.
This paper presents a computational explanation for the physiologically observed short-term synaptic depression. While most previous papers focused on its dynamic implications, this paper focuses on static pattern recognition. The paper has attracted the attention of the group of Prof. Tetsuya Asai at Hokaido University who has built a CMOS chip implementing the proposed ideas. These were further explored in a 2005 Masters project at the Swiss Federal Institute of Technology (EPFL).
This paper describes a new and robust template-based method for visually recognizing road-landmarks such as intersections. The method allows over 98% of correct execution of navigation primitives relying on vision. This is an important result showing that language-level human instructions can be converted into reliable robot programs constructed by combining high-level action primitives.
This paper reports one of the first, highly original modelling approaches in language evolution. This new methodology is based on the combination of evolutionary, neural and multi-agent computation techniques. The paper has had a significant impact in the research area of language evolution modelling, in which the author is one of the pioneer researchers. In particular it has lead establishing the importance of the process of “symbol grounding” in language studies (e.g. Steels, Kirby, Harnad). This is evidenced by its publication in one of the top ranking journals in computer science and artificial intelligence.
This paper describes work on how language influences concept formation, specifically in colour category acquisition. Theories on the nature and origins of colour categories have been controversial for several decades; we used computational modelling to contribute to the interdisciplinary discussion. In the follow-up of its publication, Belpaeme was invited to speak at several universities (including Edinburgh, Bath, UC London, Amsterdam, Gießen and Chicago). Since 2005 several researchers have adopted a similar approach to study (colour) categorisation, e.g. Jameson and Komorova (UC Irvine), Loreto (Rome) and Dowman (Tokyo).
This paper contains a detailed description of the computational model used to study how colour category typology can be explained as resulting from learning categories linguistically. The results support linguistic relativism, a theory which has for a very long been regarded as controversial. This research reaches a world-wide audience of linguists, psychologists, cognitive scientists and artificial intelligence researchers. The paper's interdisciplinary character ensures it gets considerable attention from outside computer science, and in recent years psychological research has backed up the claims laid out in the article (e.g. Gilbert, Regier, Kay & Ivry, PNAS, 2006).
This paper describes our work on building models of imitative behaviour in an attempt to build an artificial system which learns from demonstrations. There has been an increased interest in imitation, as it is “brain complete” and a very promising way of interacting with machines. This paper looks at how computers can extract the goal of a demonstration. Psychologists and computer scientists have largely ignored this problem, and only through building computational models has this been brought to the attention of an interdisciplinary audience. Both authors have held a special session on this topic at the latest IEEE Ro-Man conference.
This is an invited paper in a special issue on semi-sentient robots. It describes the representations used to convert natural language instructions into re-usable robot programs. Bugmann and colleagues were amongst the first to propose and test the IBL model (Instruction-Based Learning) for human-robot interaction. This line of work also led to invited presentations at the International Workshop on Multimodal Communication (2003, Bielefeld) and at AISB (2005, Hatfield). The journal is one of the top quality publications in the field, with impact factor consistently well above 2.
Cross-correlation is a traditional function used to analyse synchronous firings in multiple, simultaneously recorded, spike trains. This paper presents the Correlation Grid, a visualization technique that harnesses the basic function in order to make it more useful and less time consuming to analyse larger data sets. Each cross-correlogram is caricatured by a square in the Grid and then the data is processed to aid identification of clusters. This technique has been used by associates in the Laboratory of Neuro-heuristics, Grenoble. Usability testing on user groups has shown that the technique is straightforward to learn with a high degree of accuracy.
This paper challenges Pinker's hypothesis on the Baldwin Effect for the origins of the Language-specific Acquisition Device. New computer simulations on the emergence of compositional language demonstrate that Baldwinian effects are dependent on social and cognitive factors (e.g. categorical perception). These are not language-specific mechanisms, but rather general cognitive capabilities that can be later co-opted by linguistic tasks. The paper has contributed to the renewed interest in Baldwin effect studies in language evolution models and the importance of considering the interaction between evolutionary and ontogenetic factors (cf. discussion by John Holland et al. in journals Complexity and in Applied Linguistics).
This paper presents a new robotic approach to modelling the evolution of communication. It demonstrates that robots can autonomously build “linguistic” representations to communicate with each other about their sensorimotor interaction (action repertoire) with the physical environment. This research also has important implications for the development of new techniques for the design of communication capabilities in groups of autonomous agents and robots. Following this publication, Cangelosi was invited to present this approach at various conferences and meetings, including a workshop in Luxemburg to define the research priorities in Cognitive Systems and language in the EU Framework VII.
This paper presents a new developmental robotic study that models the process of first word acquisition. Its innovation is in the implementation of a learning procedure for language generativity, i.e. the generation of new concepts via language combination and grounding transfer. The simulation also demonstrates the strict relationship between action and language learning. This work also constituted the starting hypothesis for the proposal, and award, of the €6M FP7 EU project “ITALK: Integration and Transfer of Action and Language Knowledge in Robots”.
This paper presents a three-dimensional cylindrical environment for analysing the firing patterns of multiple neurons. Each of the horizontal bands, that comprise the Tunnel, encodes the spike train of the corresponding neuron. The Tunnel provides the user with a variety of frames of reference to track their location when moving or flying through the dataset. Additionally, the display is augmented with overlays that encode spike coincidence. This paper presents one of few techniques that are truly scalable for these datasets. Note that this journal is the top high-quality periodical for publication of results in the area of Information Visualization.
This paper uses the Gravity Transform (Gerstein 1985) algorithm to analyse spike train dependencies and synchronisation. Despite the pervasive use of this algorithm, the output format presented by Gerstein is both limited and complex to interpret. This paper addresses these problems by presenting a novel adaptation of the parallel coordinate display technique (Inselberg 1990) alongside dimension reduction techniques. The citations for this paper include a follow-up paper by Gerstein that incorporates an alternative output display and a review by Emery Brown in Nature Neuroscience 2004. Additionally, further work on the technique has also recently resulted in a connection with Inselberg.
This paper extends the original Harris & Wolpert optimal control model of human movement, which proposed that fast movement trajectories minimise the effects of signal-dependent noise. This paper solves the original problem analytically using optimal control theory and then shows that for saccades the observed invariant relationship between peak velocity, duration, and amplitude (‘main sequence') can be predicted as an optimal trade-off between accuracy and movement duration. The paper also considers optimal decisions of whether to move (saccade) or not (fixation). Overall, the paper demonstrates the power and parsimony of the optimality approach to understanding the underlying principles of movement.
Congenital nystagmus is a disorder of eye movement control in which the eyes oscillate incessantly from birth with very specific temporal waveforms, paradoxically often secondary to a visual deficit . This highly original paper introduces the concept that a motor disorder may develop as a plasticity strategy to maximise visual contrast when there is a loss of high spatial frequency visual information from birth. Using variational calculus, this paper shows that the predicted optimal eye movements are oscillatory and match observed waveforms. This new perspective on an age-old problem has already spawned new studies and papers.
This collaboration work between CTCN and Prof Kawato's lab (ATR Computational Neuroscience) the and Human Information Processing Research lab explores experimentally using rectified EMG signals how muscle plant dynamics change depending on the degree of co-contraction of antagonist muscles during reaching. It is found that the effects of signal-dependent noise are modulated by stiffness, such that high levels of co-contraction (increased stiffness) reduce end-point variance. High accuracy requirements induce more co-contraction and so it is argued that the nervous system also controls stiffness (impedance control) implying the existence of an additional cost parameter to the Harris & Wolpert model.
This paper challenges the assumption that human movements are ‘smooth'. By using Fourier analysis it is shown that the order of onset and offset movement discontinuities can be measured practically and that observed orders are so low that smoothness cannot be a control objective (as in the minimum jerk model). Low orders keep movement duration as low as possible and it is shown that discontinuity strengths, which can be neuromuscular or physical, constrain movement duration. This dispels a misconception about smoothness and leads to time optimality as a control objective of movement, which is taken up in output (1).
The influence of cortical feedback on sensory input to the cortex has been much debated in recent years. Here we have examined how cortical feedback mechanisms can shape the spatiotemporal responses of the cortical input signals. Our biologically-based neural model suggests that feedback plays a key role in modifying the experimentally observed reversal over time of the polarity of ON/OFF responses of the centre and surround of the spatiotemporal response. Our work was based on collaborations with Sillito and his coworkers (UCL), and their recent experimental results on cortico-thalamic feedback published in Nature Neuroscience, 2006, wherein our work is cited. Please note this publication is available on line as it is a journal publication not yet in print.
Spindles are an early sleep neuronal oscillation in the frequency range 7–14 Hz Previous theoretical studies have used detailed conductance-based models and been based on the theory that spindles are generated by intra-thalamic synaptic dynamics. Here we use a neural population based model to demonstrate that the nonlinear dynamics of the thalamocortical circuit also give rise to natural frequency oscillations in the spindle frequency range, and using bifurcation analysis we show that this oscillation is robust to biologically realistic model parameter variations. The results provide important theoretical support for the role of neural circuit dynamics in information processing.
A bidirectional associative memory (BAM), a two-layer neural network, has previously been shown to be capable of precisely learning any concept lattice structure. The objective of this paper was to show that, when the connection weights of the BAM are set according to the rule proposed by Bělohlávek to represent a specific lattice, the network always returns the most specific or most generic concept containing the given set of objects or attributes, when a set of objects or attributes is presented as input to the object or attribute layer. A proof of this property is given, together with an example.
In this paper, one of the first to apply nonlinear control theory to neuronal dynamics, neuronal membrane potential behaviour is approached as a control-theoretic “reachability” problem, in which the effect of specific stimulus-evoked synaptic inputs is to drive the neuron from an initial state to a particular terminal state on a manifold. It is shown that a fluctuating subthreshold membrane potential induced by synaptic background activity, and the cooperative action of excitatory and inhibitory inputs, may be important factors in allowing the cell to “reach” a maximal subset of possible membrane potential states, through the action of its synaptic inputs.
We review findings from a wide range of auditory streaming studies to formulate a unified account of auditory perceptual organisation, and include new computational modelling results replicating responses in primary auditory cortex, and a perceptual experiment which confirms and extends findings of bistability in auditory streaming. We propose a theoretical model of auditory perceptual organisation based on commonalities between perceptual bistability in vision and audition in which the formation of predictive models and competition between them are argued to be crucial for a system to track multiple sound sources in complex sound environments.
This paper addresses the problem of formulating a recognition model for communication sounds consisting of complex time-varying spectral patterns, for implementation in spiking neuronal networks using plausible synaptic learning mechanisms. We show that by projecting incoming activity onto a spectrotemporal feature basis set, the model can robustly classify complex sounds. The representation supports a range of qualitatively different classifications simultaneously: word, prosody, speaker identity and sex. This work led to invitations (2005, 2007) to participate as a Faculty member at the Neuromorphic Engineering Workshop, Telluride, USA, and forms the basis of a neuromorphic VLSI system being developed with ETH Zurich.
This paper describes a novel stimulus, dynamic iterated noise, which is used to demonstrate that people have a remarkable ability to perceive time varying temporal correlations in noise, as long as they vary sufficiently smoothly. These correlations are more perceptually salient than the static correlations previously used. A modified model of pitch perception which replicated the perceptual results was also presented. The predictions of this work have been tested experimentally by at least two groups, and sensitivity to time-varying temporal correlations has now been found in cochlear nucleus (Winter, Cambridge) and inferior colliculus (Gandour, Purdue).
This paper describes a 'biased competition' model of object-based visual attention that replicates human and monkey data at the cellular and systems/behavioural levels. At the cellular level, the model replicates the time-course of spatial and object-based effects in single cell data. These cellular effects lead to system level behaviour that replicates human scan paths during active visual search for a feature conjunction target. Saccade onset times and inhibition of return behaviour also mirrored that observed experimentally. The paper was reproduced in: Vision and Brain: How the Brain Sees / New Approaches to Computer Vision. Grossberg et al (eds), Elsevier, 2004
This is the first paper that suggested that climbing activity, a prominent neurophysiological phenomenon during in vivo recordings in tasks involving temporal delays, could be an indication of interval time coding allowing temporally precise predictions. A biophysically based computational model is developed that could encode arbitrary temporal intervals (across several orders of magnitude) by slope-adjustable climbing activity in line with empirical observations. In addition, the paper analyzes a novel mechanism for adapting neurons towards a line-attractor configuration (which usually requires fine-tuning) in a self-organizing manner, thereby providing a potential solution for a long-standing computational problem.
This mixed experimental and review paper discusses our computational theory of dopamine modulation of working memory. Dopamine has a profound impact on working memory, and it is shown here through state space analysis of biophysically detailed networks that dopamine modulation of synaptic currents may ultimately foster the active maintenance of working memory items and shield them from distraction. The paper demonstrates how in vitro electrophysiological findings at a biophysical level can be functionally interpreted and linked to in vivo electrophysiology and cognition by computational modeling. With nearly 50 citations the paper had quite some impact on the field.
This is the first paper to clearly demonstrate that dopamine, a substance of major importance in learning and working memory, has both time-dependent and opposing effects on inhibitory currents via D1- and D2-receptors. With over 100 citations the paper already had a major impact on the field, and laid the groundwork for our current influential computational theory of dopamine modulation which poses that dopamine may push cortical networks either into a D1-dominated regime highly beneficial for robust maintenance of working memory items, or a D2-dominated regime allowing high cognitive flexibility at the expense of working memory.
Cortical self-organisation is believed to follow optimisation principles like the maximisation of information transmitted by neurons (Infomax). This work shows that spike-timing-dependent plasticity (STDP) supports the maximisation of temporal stochastic interactions in recurrent neural networks and induces automata- or rule-like behaviour in such systems. It is one of only very few studies that consider recurrent networks and spatio-temporal correlations in contrast to just feedforward systems and one-time correlations. Several robotics groups (O.Sporns, Indiana University, USA; D.Polani, University of Hatfield, UK; R.Der, University of Leipzig, Germany) are now adopting comparable approaches for self-organising behaviour in embodied robots. Our framework underpins information-theoretic aspects of these approaches.
This work presents an approximation method for analysing dynamic receptive field properties as they are inceasingly observed in sensory systems of mammals. It provides a general mathematical method for the analyis of layered extended models for cortical maps in the cortex which can be adapated to a large number of experimental paradigms using localised stimuli in the visual, auditory, or somato-sensory domain. It provides simplified descriptions of the occurring phenomena, analytical solutions in special cases, and an intuitive understanding of which dynamic processes shape cortical tuning.
Simple cells in primary visual cortex often show constant orientation tuning but a nonlinear amplitude response with contrast. A number of international groups have explored this behaviour in theoretical and simulation studies (e.g., Somers et al. 1995, Ben Yishay et al. 1995, Carandini and Ringach 1997). This paper presents a fully non-linear analytical study of simple cells which provides explicit equations for tuning properties observed in the previous, mostly computational works. The approach is in addition much more general than the earlier ones. It also reconciles these models with the apparently conflicting one by Troyer et al. (1998) by interpolating smoothly between both extremes.
This paper presents a neural network model which is able to track multiple objects. The model is based on the principle of synchronization of neural activity and the focus of attention is described in terms of resonant oscillators. The model is deeply grounded in experimental data on gamma rhythm and synchronization modulated neural activity. The model demonstrates features which are similar to findings of psychological experiments. In particular, the curve of the model errors versus the number of tracked objects is close to experimental measurements.
This paper describes a mathematical framework for understanding the basic mechanisms of memory and learning. Current experimental evidence from neurobiology on neural, synaptic, and molecular mechanisms of memory are as yet incomplete and quite contradictory. They do not provide a basis for a theoretical study of the learning process. This paper provides a general theory of memory and learning which both incorporates general neurobiological principles and is open to future experimental investigations.
This paper aims at demonstrating a hierarchical neural architecture to execute several cognitive functions: selective attention, object memorization and novelty detection. This model is one of only a few which aim at and successfully model several cognitive functions by the same neural architecture.
This paper describes a variant of the enhanced integrate and fire model which includes some important features of real neurons and significantly advances the traditional integrate and fire based approach to modelling of single neural activity. The study of the dynamical regimes of a network of excitatory and inhibitory elements for the first time reveals two different scenarios of appearance of oscillations which are similar to the standard Andronov-Hopf and saddle-node invariant curve bifurcations.
This paper discusses the conditions in which artificial agents must be able to integrate sensory-motor information over time and later use this information to modulate their behaviour. In particular, the focus of the paper is on the transition from simple agents, that only rely on sensory information or on their internal dynamic, to agents that are also able to rely on mixed strategy, in which basic sensory-motor mechanisms are complemented and enhanced with additional internal mechanisms. Moreover, we show that evolved individuals tend to rely on partial, action-oriented, and action-mediated representations of the external environment.
This paper is the first clear demonstration that dopamine through D1 receptors may have differential effects on slow NMDA vs. fast AMPA excitatory synaptic current. This differential modulation turned out to be computationally highly important in network simulations as it provides one basis for the dopaminergic enhancement of persistent (working memory-related) activity but suppression of brief (potentially distracting) inputs.
Persistent activity is assumed to be the neural hallmark of the short-term maintenance of stimulus- or goal-related information in prefrontal cortex during working memory but it's underlying synaptic mechanism are only poorly understood. In this paper we demonstrate that the slow drive provided by recurrent NMDA synaptic inputs may be crucial for the maintenance of high activity states while AMPA currents may play a role in initiating such states.
This paper was highly important because it not only revealed a novel signaling cascade for dopamine receptors, it also provided a mechanism for how different dopamine receptors subtypes (D1 vs. D2) may become selectively activated. This has been a puzzling question for some time as dopamine activates both D1- and D2-receptors in prefrontal cortex, yet they have often opposing effects. Resolving this issue was therefore an important step in understanding the computational function of dopamine in prefrontal cortex.
This paper reports on an initial model using an agent-based methodology to study the mutual impact of language on colour categorisation. It reports pilot experiments on how colour typology, as observed in various psychological experiments, can be explained as resulting from dynamics of individuals learning and using colour terms. The paper was orally presented at IJCAI2001, a large and competitive (25% acceptance rate in 2001) conference on artificial intelligence. The work described in this paper evolved in studies subsequently reported in two journal publications.

This is a computational modeling study of experiments previously described in a Nature paper by some of the authors (Worgotter et al., Nature 396: 165–168, 1998).

It is one of the first modeling studies that considered dynamic receptive field effects in primary visual cortex, in particular receptive field sharpening on short time scales (ca 100ms) and its dependence on the behavioural state (weak sleep, aroused state) of the animal. The computational modeling in the paper is further corroborated by a rigorous mathematical analysis of a simplified abstract model for the studied phenomena. As such the work provides an early and strong argument for a dynamic adaption of sensory processing on short time scales.

The paper shows that complex visual tasks, such as position- and size-invariant shape recognition and navigation, can be tackled with simple architectures generated by a co-evolutionary process of active vision and feature selection. We show that robots autonomously develop sensitivity to a number of emergent visual features such as retinotopic visual-feature-oriented edges, corners, and height. Robots also develop a behavioural repertoire to locate, bring, and keep these features in sensitive regions of the vision system, resembling strategies observed in simple insects. This work resulted from a research visit at the Swiss Federal Institute of Technology (EPFL).
This paper presents new evolutionary robotic experiments in which a collection of simulated robots, that have been evolved for the ability to solve a collective navigation problem, also develop a communication system that allows them to co-operate better. The lexicon analysis demonstrates that robots develop a non-trivial communication system and exploit different adaptive communication modalities. The results also indicate how the possibility of co-adapting the robots' individual and social/communicative behaviour plays a key role in the development of progressively more complex and effective artificial agents.
This paper proposes the adaptation and further development of the Synthetic Brain Image methodology for cognitive neural network models. The novel application of such a technique to an evolutionary model of language learning demonstrates that it is possible to constrain neural network topologies that (qualitatively) reflect knows neuroimaging phenomena (e.g. differentiation of areas for the processing of verbs and of nouns). The first author has been invited to present such a work at international workshops such as the ESF workshop on Neurobiology of Communication (Cambridge, 2002) and to successfully submit it to one of the main journals in language neuroscience.
We had suggested in Seamans & Yang (2004) that dopamine acts on different time courses through different signaling mechanisms. This paper employed a variety of sophisticated techniques to show that the dopamine system could signal fast events via co-release of glutamate while modulating intracortical processing on protracted timescales via dopamine release. This was a novel perspective that potentially solved a number of long-standing issues in the field. In particular, it provides a potential resolution for the dual function of dopaminergic neurons in signaling prediction errors as proposed by Schultz and Dayan on the one hand side, while modulating prefrontal cortex attractor states in line with the computational theory proposed by Durstewitz and Seamans on the other hand.