Search

There is a growing requirement in computational neuroscience for tools that permit collaborative model building, model sharing, combining existing models into a larger system (multi-scale model integration), and are able to simulate models using a variety of simulation engines and hardware platforms. Layered XML model specification formats solve many of these problems, however they are difficult to write and visualise without tools. Here we describe a new graphical software tool, SpineCreator, which facilitates the creation and visualisation of layered models of point spiking neurons or rate coded neurons without requiring the need for programming. We demonstrate the tool through the reproduction and visualisation of published models and show simulation results using code generation interfaced directly into SpineCreator. As a unique application for the graphical creation of neural networks, SpineCreator represents an important step forward for neuronal modelling.

Decision formation recruits many brain regions, but the procedure they jointly execute is unknown. Here we characterize its essential composition, using as a framework a novel recursive Bayesian algorithm that makes decisions based on spike-trains with the statistics of those in sensory cortex (MT). Using it to simulate the random-dot-motion task, we demonstrate it quantitatively replicates the choice behaviour of monkeys, whilst predicting losses of otherwise usable information from MT. Its architecture maps to the recurrent cortico-basal-ganglia-thalamo-cortical loops, whose components are all implicated in decision-making. We show that the dynamics of its mapped computations match those of neural activity in the sensorimotor cortex and striatum during decisions, and forecast those of basal ganglia output and thalamus. This also predicts which aspects of neural dynamics are and are not part of inference. Our single-equation algorithm is probabilistic, distributed, recursive, and parallel. Its success at capturing anatomy, behaviour, and electrophysiology suggests that the mechanism implemented by the brain has these same characteristics., Author summary Decision-making is central to cognition. Abnormally-formed decisions characterize disorders like over-eating, Parkinson’s and Huntington’s diseases, OCD, addiction, and compulsive gambling. Yet, a unified account of decision-making has, hitherto, remained elusive. Here we show the essential composition of the brain’s decision mechanism by matching experimental data from monkeys making decisions, to the knowable function of a novel statistical inference algorithm. Our algorithm maps onto the large-scale architecture of decision circuits in the primate brain, replicating the monkeys’ choice behaviour and the dynamics of the neural activity that accompany it. Validated in this way, our algorithm establishes a basic framework for understanding the mechanistic ingredients of decision-making in the brain, and thereby, a basic platform for understanding how pathologies arise from abnormal function.

Small Unmanned Air Vehicles (UAVs) have become increasingly used in many fields to keep up with demands for technological growth and due to their unique ability to provide an “eye-in-the-sky”. However, traditional guidance systems are not always suitable for a transition to smaller UAVs. Honeybee navigation has long been proposed as a basis for developing navigation for robotics as they are known to solve complex tasks efficiently and robustly. An approach for bio-inspired navigation using optic flow is presented here based on honeybee reactive flight control. This approach is tested and verified in the benchmark hallway navigation problem. It is shown that the control approach can explain a wide-range of biological behaviors using minimal sensors similar to flying insects.

The basal ganglia are a subcortical group of interconnected nuclei involved in mediating action selection within cortex. A recent proposal is that this selection leads to optimal decision making over multiple alternatives because the basal ganglia anatomy maps onto a network implementation of an optimal statistical method for hypothesis testing, assuming that cortical activity encodes evidence for constrained gaussian-distributed alternatives. This letter demonstrates that this model of the basal ganglia extends naturally to encompass general Bayesian sequential analysis over arbitrary probability distributions, which raises the proposal to a practically realizable theory over generic perceptual hypotheses. We also show that the evidence in this model can represent either log likelihoods, log-likelihood ratios, or log odds, all leading proposals for the cortical processing of sensory data. For these reasons, we claim that the basal ganglia optimize decision making over general perceptual hypotheses represented in cortex. The relation of this theory to cortical encoding, cortico-basal ganglia anatomy, and reinforcement learning is discussed.

Izhikevich (2003) proposed a new canonical neuron model of spike generation. The model was surprisingly simple yet able to accurately replicate the firing patterns of different types of cortical cell. Here, we derive a solution method that allows efficient simulation of the model.

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.

Neural networks are modelling tools that are, in principle, able to capture the input–output behaviour of arbitrary systems that may include the dynamics of animal populations or brain circuits. While a neural network model is useful if it captures phenomenologically the behaviour of the target system in this way, its utility is amplified if key mechanisms of the model can be discovered, and identified with those of the underlying system. In this review, we first describe, at a fairly high level with minimal mathematics, some of the tools used in constructing neural network models. We then go on to discuss the implications of network models for our understanding of the system they are supposed to describe, paying special attention to those models that deal with neural circuits and brain systems. We propose that neural nets are useful for brain modelling if they are viewed in a wider computational framework originally devised by Marr. Here, neural networks are viewed as an intermediate mechanistic abstraction between ‘algorithm’ and ‘implementation’, which can provide insights into biological neural representations and their putative supporting architectures.

The existence of multiple parallel loops connecting sensorimotor systems to the basal ganglia has given rise to proposals that these nuclei serve as a selection mechanism resolving competitions between the alternative actions available in a given context. A strong test of this hypothesis is to require a computational model of the basal ganglia to generate integrated selection sequences in an autonomous agent, we therefore describe a robot architecture into which such a model is embedded, and require it to control action selection in a robotic task inspired by animal observations. Our results demonstrate effective action selection by the embedded model under a wide range of sensory and motivational conditions. When confronted with multiple, high salience alternatives, the robot also exhibits forms of behavioral disintegration that show similarities to animal behavior in conflict situations. The model is shown to cast light on recent neurobiological findings concerning behavioral switching and sequencing.

We present a novel deterministic search technique for finding a set of maximal conductances in multi-compartmental models, and apply this to a model of a striatal medium spiny neuron. The search rapidly converges on a set of maximal conductances by fitting simulation output to current injection data. The search is robust, finds solutions in the presence of a multitude of inward and outward ionic currents, and is extendible to include a subset of the passive membrane parameters. After incorporating synaptic input, the resulting model of the medium spiny neuron provides a good fit to a range of experimental data.

The field of autonomous vehicle control (AVC) is a rapidly growing one which promises improved performance, autonomy, comfort and safety for future’s intelligent transportation systems (ITS). In this paper, the objective is to present and demonstrate a cognitive modular approach for adaptive cruise control of autonomous/driverless vehicles by exploiting similarities between signal processing mechanisms and system architectures in control systems and the animal brain. To demonstrate the effectiveness of the proposed cognitive cruise control approach, numerical results from a case study with a realistic vehicle model is provided in this paper.

We address the question of how the visual system analyses changes in direction. Using plaid stimuli, we define type O direction changes which entail a change in the orientations of the plaid components, and type V direction changes in which the orientations of the components remain constant, relative to the observer but their relative speeds change. Lower thresholds for discriminating type O and type V direction changes were compared. Type O thresholds for clockwise/anticlockwise direction change were very low (0.2-0.5 degree), were resistant to directional noise, and showed a low-pass relationship with drift velocity. Type V thresholds on the other hand were higher (1-5 degrees), and exhibited a bandpass relationship with drift velocity. Type O direction changes gave low thresholds at short inter-stimulus intervals (ISI) (< 160 ms) and higher thresholds (successive orientation discrimination) at long ISI (240 ms-12.8 s). Type V thresholds, on the other hand, exhibited no short-range process and performance at short ISI, was no better than for successive direction discrimination at long ISI. A two-stage rotary motion model is sufficient to explain the discrimination of type O direction changes and results rule out a model based on velocity discrimination. For type V direction changes, a two-stage mechanism is insufficient and results are consistent with a minimum of three computational stages.

We revieiv recent research in robotics, neuroscience, evolutionary neurobiology, and ethology with the aim of highlighting some points of agreement and convergence. Specifically, we com pare Brooks' (1986) subsumption architecture for robot control with research in neuroscience demonstrating layered control systems in vertebrate brains, and with research in ethology that emphasizes the decomposition of control into multiple, intertwined behavior systems. From this perspective we then describe interesting parallels between the subsumption architecture and the natural layered behavior system that determines defense reactions in the rat. We then consider the action selection problem for robots and vertebrates and argue that, in addition to subsumption- like conflict resolution mechanisms, the vertebrate nervous system employs specialized selection mechanisms located in a group of central brain structures termed the basal ganglia. We suggest that similar specialized switching mechanisms might be employed in layered robot control archi tectures to provide effective and flexible action selection.

Intrinsic motivations drive an agent to explore, providing essential data for linking behaviors with novel outcomes and so laying the foundation for future flexible action. We present experiments using a new behavioral task which allows us to interrogate the connection between exploration and action learning. Human participants used a joystick to search repeatedly for a target location, only receiving feedback on successful discovery. Feedback delay was manipulated, as was the starting position. Experiment 1 employed stable starting positions, so the task could be learnt with respect to a target location or a target trajectory. Participants were able to learn the correct movement under all delay conditions. Experiment 2 used a variable starting location, so the correct movement could only be learnt in terms of target location. Participants displayed little to no learning in this experiment. These results suggest that movements on this scale are stored as trajectories rather than in terms of target location. Overall the experiments demonstrate the potential of this task for uncovering the native representational substrates of action learning.

The development of an 'artificial brain' is one of the greatest challenges in artificial intelligence, and its success will have innumerable benefits in many and diverse fields, including the creation of autonomous robotic agents. Most research effort is spent on modelling vertebrate brains. Yet smaller brains can display comparable cognitive sophistication, while being more experimentally accessible and amenable to modelling.

Reinforcement (trial-and-error) learning in animals is driven by a multitude of processes. Most animals have evolved several sophisticated systems of 'extrinsic motivations' (EMs) that guide them to acquire behaviours allowing them to maintain their bodies, defend against threat, and reproduce. Animals have also evolved various systems of 'intrinsic motivations' (IMs) that allow them to acquire actions in the absence of extrinsic rewards. These actions are used later to pursue such rewards when they become available. Intrinsic motivations have been studied in Psychology for many decades and their biological substrates are now being elucidated by neuroscientists. In the last two decades, investigators in computational modelling, robotics and machine learning have proposed various mechanisms that capture certain aspects of IMs. However, we still lack models of IMs that attempt to integrate all key aspects of intrinsically motivated learning and behaviour while taking into account the relevant neurobiological constraints. This paper proposes a bio-constrained system-level model that contributes a major step towards this integration. The model focusses on three processes related to IMs and on the neural mechanisms underlying them: (a) the acquisition of action-outcome associations (internal models of the agent-environment interaction) driven by phasic dopamine signals caused by sudden, unexpected changes in the environment; (b) the transient focussing of visual gaze and actions on salient portions of the environment; (c) the subsequent recall of actions to pursue extrinsic rewards based on goal-directed reactivation of the representations of their outcomes. The tests of the model, including a series of selective lesions, show how the focussing processes lead to a faster learning of action-outcome associations, and how these associations can be recruited for accomplishing goal-directed behaviours. The model, together with the background knowledge reviewed in the paper, represents a framework that can be used to guide the design and interpretation of empirical experiments on IMs, and to computationally validate and further develop theories on them.

A model of local image encoding is described which explicitly incorporates quantitative data about the number density, bandwidth and receptive field organisation of neurons involved in motion detection. The model solves the problem of extracting local velocity on the basis of inputs tuned to spatiotemporal frequency and sensitive to contrast. The spatiotemporally tuned, opponent motion filters are followed by a compressive non-linearity and comprise a first stage. The inter-stage signals are interpreted as those from single neurons and the second stage is modelled as a neural-network layer. The second stage uses semilinear units and models the effect of lateral, on-centre off-surround, intra-layer connections. Characterisation of the first stage leads to a clarification of the concept of the psychophysical 'channel' and its relation to physiological data. The quantitative parametrisation of the model allows the simulation of several psychophysical phenomena which are reported in a companion paper.

Lower motion thresholds for rotational and radial flow have been measured for stimuli consisting of four closely packed circular apertures, each containing patches of drifting grating or plaid. Detection and direction thresholds were measured for gratings and plaids as a function of the relative orientation of the pattern components. There was a similarity between both types of threshold, supporting the existence of specialised rotation and radial-flow detectors. Further, thresholds increased with the relative component orientation for both gratings and plaids. This suggests that component information from a first stage, tuned spatiotemporally and to orientation, is being used directly to compute the optic flow in a two-stage process. A model based on this architecture is described by means of simple template receptive-field arrays with separable temporal and spatial tuning at the first stage.

We present a novel behavioural task for the investigation of how actions are added to an agent’s repertoire. In the task, free exploration of the range of possible movements with a manipulandum, such as a joystick, is recorded. A subset of these movements trigger a reinforcing signal. Our interest is in how those elements of total behaviour which cause an unexpected outcome are identified and stored. This process is necessarily prior to the attachment of value to different actions [Redgrave, P., Gurney, K.: The short-latency dopamine signal: A role in discovering novel actions? Nat. Rev. Neurosci. 7(12), 967–975 (2006)]. The task allows for critical tests of reinforcement prediction error theories [e.g. Schultz, W., Dayan, P., Montague, P.: A neural substrate of prediction and reward. Science 275, 1593–1599 (1997)], as well as providing a window on a number of other issues in action learning. The task provides a paradigm where the exploratory motive drives learning, and as such we view it as in the tradition of Thorndike [Animal intelligence (1911)]. Our task is easily scalable in difficulty, is adaptable across species and provides rich set of behavioural measures throughout the action-learning process. Targets can be defined in spatial, temporal or kinematic/gestural terms, and the task also allows the concatenation of actions to be investigated. Action learning requires integration across spatial, kinematic and temporal dimensions. The task affords insight into these (and into the process of integration).

We introduce a biologically motivated, formal framework or “ontology” for dealing with many aspects of action discovery which we argue is an example of intrinsically motivated behaviour (as such, this chapter is a companion to that by Redgrave et al. in this volume). We argue that action discovery requires an interplay between separate internal forward models of prediction and inverse models mapping outcomes to actions. The process of learning actions is driven by transient changes in the animal’s policy (repetition bias) which is, in turn, a result of unpredicted, phasic sensory information (“surprise”). The notion of salience as value is introduced and broken down into contributions from novelty (or surprise), immediate reward acquisition, or general task/goal attainment. Many other aspects of biological action discovery emerge naturally in our framework which aims to guide future modelling efforts in this domain.

Studies of decision making in animals suggest a neural mechanism of evidence accumulation for competing percepts according to Bayesian sequential analysis. This model of perception is embodied here in a biomimetic tactile sensing robot based on the rodent whisker system. We implement simultaneous perception of object shape and location using two psychological test paradigms: first, a free-response paradigm in which the agent decides when to respond, implemented with Bayesian sequential analysis; and second an interrogative paradigm in which the agent responds after a fixed interval, implemented with maximum likelihood estimation. A benefit of free-response Bayesian perception is that it allows tuning of reaction speed against accuracy. In addition, we find that large gains in decision performance are achieved with unforced responses that allow null decisions on ambiguous data. Therefore free-response Bayesian perception offers benefits for artificial systems that make them more animal-like in behavior.

Texture perception is studied here in a physical model of the rat whisker system consisting of a robot equipped with a biomimetic vibrissal sensor. Investigations of whisker motion in rodents have led to several explanations for texture discrimination, such as resonance or stick-slips. Meanwhile, electrophysiological studies of decision-making in monkeys have suggested a neural mechanism of evidence accumulation to threshold for competing percepts, described by a probabilistic model of Bayesian sequential analysis. For our robot whisker data, we find that variable reaction-time decision-making with sequential analysis performs better than the fixed response-time maximum-likelihood estimation. These probabilistic classifiers also use whatever available features of the whisker signals aid the discrimination, giving improved performance over a single-feature strategy, such as matching the peak power spectra of whisker vibrations. These results cast new light on how the various proposals for texture discrimination in rodents depend on the whisker contact mechanics and suggest the possibility of a common account of decision-making across mammalian species.

Redgrave and Gurney [1] have recently proposed a theory of action discovery in which the occurrence of an unexpected salient sensory event causes short-latency phasic dopamine (DA) neuron activity. Due to convergent projections from DA neurons and cortical neurons onto the striatum, the DA signal reinforces (i.e., biases the animal to repeat) preceding behaviour. In the case that the animal has caused the sensory outcome, continued repetition, along with variations in behaviour, allows the animal to discover the novel action-outcome contingency. A key aspect of this theory is that the DA signal decreases as the action-outcome association becomes predictable. In contrast is the well-known theory that the DA signal specifically represents a reward prediction error [2], and that the tendency to select behaviour is based on its ‘value’ (often represented as the expected subsequent cumulative costs and rewards). The latter theory is attractive in part because it conforms with Reinforcement Learning (RL), a computational formulation of learning from the consequences of actions [3]. RL describes well behaviour observed during many types of goal-directed tasks and allows behaviour to be described as optimal given a particular reward/cost function (e.g., cost per movement). However, is a notion of optimality necessary to produce all behaviour thought to derive from DA-mediated mechanisms? If so, a representation of every movement’s cost and value is required to govern learning. The theory of action discovery [1] suggests that some behaviour is developed through much simpler mechanisms: the occurrence of the salient event rather than a prediction of some quantity (value) associated with each action selected. Here, we implement such a scheme in a learning agent that chooses discrete actions within a discrete state environment (analogous to pushing buttons). Only when the agent has reached a particular state (the salient event) is a reinforcement signal delivered. The tendency to select every action that was selected en route to that state is increased, but with a learning rate that decreases with temporal distance from the reinforcement signal (i.e., an eligibility trace). Importantly, no notion of optimality is represented. Figure ​Figure11 below illustrates the development of behavior due to our scheme (black line) and an RL scheme based on cost (grey). In both, the number of actions chosen en route to the salient event decreases toward a minimum. Such behavior is usually taken as evidence that it approaches optimal given some cost function on which learning depends. However, with our scheme, such behavior emerges ‘for free’ from the statistics of behavior over repeated trials. Actions that lie along the most direct path simply get reinforced at a greater rate than other actions. Figure 1 If, as suggested by [1], learning is truncated after the outcome becomes predictable (via gradual extinction of phasic DA neuron activity), behavior under both schemes will stabilize to the most direct path. However, our scheme also predicts that, without such a truncation, prolonged ‘overlearning’ will result in suboptimal behavior where there is an increasing tendency to select spurious actions (see figure ​figure1).1). Such behavior may be analogous to behavior arising from certain types of DA dysfunction. We discuss the implications of this learning scheme within the context of action discovery and describe additional predicted behaviors.

Previously, it has been shown experimentally that the psychophysical law known as Pieron’s Law holds for color intensity and that the size of the effect is additive with that of Stroop condition (Stafford et al., 2011). According to the additive factors method (Donders, 1868–1869/1969; Sternberg, 1998), additivity is assumed to indicate independent and discrete processing stages. We present computational modeling work, using an existing Parallel Distributed Processing model of the Stroop task (Cohen et al., 1990) and a standard model of decision making (Ratcliff, 1978). This demonstrates that additive factors can be successfully accounted for by existing single stage models of the Stroop effect. Consequently, it is not valid to infer either discrete stages or separate loci of effects from additive factors. Further, our modeling work suggests that information binding may be a more important architectural property for producing additive factors than discrete stages.

A general problem in robotics is how to best utilize sensors to classify the robot's environment. The BIOTACT project (BIOmimetic Technology for vibrissal Active Touch) is a collaboration between biologists and engineers that has led to many distinctive robots with artificial whisker sensing capabilities. One problem is to construct classifiers that can recognize a wide range of whisker sensations rather than constructing different classifiers for specific features. In this article, we demonstrate that a stationary naive Bayes classifier can perform such a general classification by applying it to various robot experiments. This classifier could be a key component of a robot able to learn autonomously about novel environments, where classifier properties are not known in advance.

Navigating the visual world is a challenging problem for autonomous agents, which must be flexible, robust, and preferably easily extensible in order to meet changing task demands. Here, we outline the rationale for an approach to constructing such agents biomimetically, even if they appear ‘over engineered’ at first glance, using the problem of gaze redirection in an attentional task.

Novelty detection would be a useful ability for any autonomous robot that seeks to categorize a new environment or notice unexpected changes in its present one. A biomimetic robot (SCRATCHbot) inspired by the rat whisker system was here used to examine the performance of a novelty detection algorithm based on a “naive” implementation of Bayes rule. Naive Bayes algorithms are known to be both efficient and effective, and also have links with proposed neural mechanisms for decision making. To examine novelty detection, the robot first used its whiskers to sense an empty floor, after which it was tested with a textured strip placed in its path. Given only its experience of the familiar situation, the robot was able to distinguish the novel event and localize it in time. Performance increased with the number of whiskers, indicating benefits from integrating over multiple streams of information. Considering the generality of the algorithm, we suggest that such novelty detection could have widespread applicability as a trigger to react to important features in the robot's environment.

Action selection in animals requires rapid decision making that can discriminate the most salient requests for behavioral expression. The basal ganglia (BG) are believed to play a critical role in resolving competition between these requests [1] and, recently it has been proposed that the BG and cortex, taken together, implement a decision algorithm known as the multi-hypothesis sequential probability ratio test (MSPRT) [2]. Here, the cortex first integrates noisy ‘evidence’ indicating salience of action requests. The BG then examine this integrated evidence, and report the channel with maximal mean salience. The MSPRT is optimal in the sense that it guarantees the smallest decision time for a given error. The signals used in [2] assume cortical evidence - as neural firing rates - is supplied by Gaussian distributed signals. However, it is known that such rates are not distributed normally. To remedy this we have deduced three new forms for MSPRT that are based on more biologically realistic input signals. They include the Inverse Gaussian, Gamma and Lognormal distributions. All implementations preserve a form which could be mapped, in principle, onto the anatomy of the BG. Figure ​Figure11 shows an example of the performance of the MSPRT models across all distributions, together with the so called race model. All variants used the same input statistics (mean and variance). Counterintuitively, the relative performance of the models did not change with base level of salience or evidence. Figure 1 Performance of original Gaussian MSPRT[2], new implementations of MSPRT, and Race model. Performance is measured as decision time against number of alternatives for selection. The selection criterion was 1% error rate. The results are averaged across ... A critical parameter in the algorithm is the time step Δt used to sample the cortical input. The Gaussian, Inverse Gaussian, and Gamma cases are generated as stable Levy processes and may be scaled indefinitely with decreasing Δt. However, unlike the MSPRT with Gaussian signals, we show that decision time for the inverse Gaussian, and Gamma decreases with Δt. Figure ​Figure11 use a Δt = 1 ms for all models and clearly shows a performance advantage for the cases with more realistic distributions (the lognormal case is shown with similar scaling even though it has not an associated Levy process). Indefinitely small decision times may appear unrealistic, but we present an interpretation of cortical sampling that relates Δt to the inter-spike intervals of an ensemble of neural afferents impinging on the cortical ‘integrators’. This and, other mechanisms, provide a natural lower bound for Δt. We conclude that neurons may take advantage of the properties of natural spike trains to enhance decision making in the cortex-BG complex.

Biological computational modellers are becoming increasingly interested in building large, eclectic models, including components on many different computational substrates, both biological and non-biological. At the same time, the rise of the philosophy of embodied modelling is generating a need to deploy biological models as controllers for robots in real-world environments. Finally, robotics engineers are beginning to find value in seconding biomimetic control strategies for use on practical robots. Together with the ubiquitous desire to make good on past software development effort, these trends are throwing up new challenges of intellectual and technological integration (for example across scales, across disciplines, and even across time) - challenges that are unmet by existing software frameworks. Here, we outline these challenges in detail, and go on to describe a newly developed software framework, BRAHMS. that meets them. BRAHMS is a tool for integrating computational process modules into a viable, computable system: its generality and flexibility facilitate integration across barriers, such as those described above, in a coherent and effective way. We go on to describe several cases where BRAHMS has been successfully deployed in practical situations. We also show excellent performance in comparison with a monolithic development approach. Additional benefits of developing in the framework include source code self-documentation, automatic coarse-grained parallelisation, cross-language integration, data logging, performance monitoring, and will include dynamic load-balancing and 'pause and continue' execution. BRAHMS is built on the nascent, and similarly general purpose, model markup language, SystemML. This will, in future, also facilitate repeatability and accountability (same answers ten years from now), transparent automatic software distribution, and interfacing with other SystemML tools. (C) 2009 Elsevier Ltd. All rights reserved.

Many rodents use their whiskers to distinguish objects by surface texture. To examine possible mechanisms for this discrimination, data from an artificial whisker attached to a moving robot was used to test texture classification algorithms. This data was examined previously using a template-based classifier of the whisker vibration power spectrum [1]. Motivated by a proposal about the neural computations underlying sensory decision making [2], we classified the raw whisker signal using the related ‘naive Bayes’ method. The integration time window is important, with roughly 100ms of data required for good decisions and 500ms for the best decisions. For stereotyped motion, the classifier achieved hit rates of about 80% using a single (horizontal or vertical) stream of vibration data and 90% using both streams. Similar hit rates were achieved on natural data, apart from a single case in which the performance was only about 55%. Therefore this application of naive Bayes represents a biologically motivated algorithm that can perform well in a real-world robot task.

We present a biologically inspired neural network model of visual orienting (using saccadic eye movements) in which targets are preferentially selected according to their reward value. Internal representations of visual features that guide saccades are developed in a self-organised map whose plasticity is modulated under reward. In this way, only those features relevant for acquiring rewarding targets are generated. As well as guiding the formation of feature representations, rewarding stimuli are stored in a working memory and bias future saccade generation. In addition, a reward prediction error is used to initiate retraining of the self-organised map to generate more efficient representations of the features when necessary.

The dynamics and training of recurrent nets which make use of probabilistic nodes based on Boolean functions are explored. These are shown to be equivalent to sigma-pi units when using binary inputs and have the advantage of an immediate implementation in readily available digital hardware. The underlying hypercube structure of the Boolean address space is used to describe the process of generalisation and also sheds light on the role of hidden units. This is made more precise with the aid of the concept of hypercube order statistics. Training algorithms are described which make use of a node time-integration apparatus to promote near optimal cube structures in a natural way.

Lower motion thresholds for discriminating opposing motion directions were compared for one dimensional (grating) and two dimensional (plaid) stimuli in central and peripheral vision. The results were consistent with a two-stage model of motion sensitivity in which threshold-limiting noise occurs at both stages, and the speed as well as the direction of the resultant motion is determined by intersection-of-constraints (IOC) from the component motions. The results do not support a purely geometric interpretation of the IOC model, in which thresholds for plaid stimuli are related to thresholds of component gratings by a geometric factor. Neither do the data favour explanations in which local luminance features (i.e. blobs) are detected and their velocity determined. Monte-Carlo simulations of the two-stage process predict thresholds across variations in component direction, contrast, and visual field eccentricity. Lower motion thresholds for gratings and plaids both follow a saturating function of contrast; the fit between grating and plaid data is improved when the plaid contrast is expressed in terms of the contrast of its components. Although less contrast saturation was found in the periphery, in relative terms, plaid and grating motion thresholds were similar in central and peripheral vision, implying cortical magnifications are similar for mechanisms which process grating and plaid motion.

INTRODUCTION In the visual system, attending to important objects in the visual field relies on the transfer of top-down, object-based task information to the spatially organised areas of cortex. How this occurs and the method by which this information can influence the dorsal stream and redirect gaze are not well understood. Current models of the ventral stream mostly focus on the feed-forward mechanisms involved and current feedback models do not seem to address the issue of object-space binding in a comprehensive and plausible manner. METHODS We investigated these questions using the following modeling framework. A bidirectional, ventral stream object recognition hierarchy up to anterior inferior temporal cortex (AIT) from primary visual cortex (V1) and a model of dorsal stream to frontal eye fields (FEF) with our previously developed oculomotor system [1]. Selection is performed in both the object-based mapping of AIT [2] and the spatial mapping of FEF [3] by basal ganglia loops [4]. Modeling of the ventral stream consists of a hierarchy of increasingly spatially invariant cortical areas linked by both feed-forward excitatory and feedback connections. Within each receptive field, there is a competition to represent the strongest and thus most likely representation for that region, which can be biased by the feedback from higher visual areas. Three models of feedback attention mechanism were tested: additive feedback, shunting (multiplicative) feedback and a shunt "gating" of feedback by feed-forward. The model was tested using a simple visual world (colored "flags") that nevertheless challenged all the main competencies being investigated. Performance was measured by (i) eliciting saccadic "behavior" in simulated visual search with different numbers of distractors, and (ii) target segmentation in cluttered scenes within a fixed time window. RESULTS In the target segmentation task, the additive feedback model consistently fails to bind the AIT representation of the object to the correct location on the visual field. The shunting model was able to segment 58% of scenes while the gating model was most successful (83%) (Figure 1). We then took the most successful (gating) model and challenged it with a conjunction visual search task. Here, by simulating models trained and naive to the target stimulus, we showed that subsequent learning of a combined representation of an untrained target stimulus can explain the experimentally observed decrease in the slope of reaction time against number of distractors for that target (Figure 2) [5].

The striatum, the principal input structure of the basal ganglia, is crucial to both motor control and learning. It receives convergent input from all over the neocortex, hippocampal formation, amygdala and thalamus, and is the primary recipient of dopamine in the brain. Within the striatum is a GABAergic microcircuit that acts upon these inputs, formed by the dominant medium-spiny projection neurons (MSNs) and fast-spiking interneurons (FSIs). There has been little progress in understanding the computations it performs, hampered by the non-laminar structure that prevents identification of a repeating canonical microcircuit. We here begin the identification of potential dynamically-defined computational elements within the striatum. We construct a new three-dimensional model of the striatal microcircuit's connectivity, and instantiate this with our dopamine-modulated neuron models of the MSNs and FSIs. A new model of gap junctions between the FSIs is introduced and tuned to experimental data. We introduce a novel multiple spike-train analysis method, and apply this to the outputs of the model to find groups of synchronised neurons at multiple time-scales. We find that, with realistic in vivo background input, small assemblies of synchronised MSNs spontaneously appear, consistent with experimental observations, and that the number of assemblies and the time-scale of synchronisation is strongly dependent on the simulated concentration of dopamine. We also show that feed-forward inhibition from the FSIs counter-intuitively increases the firing rate of the MSNs. Such small cell assemblies forming spontaneously only in the absence of dopamine may contribute to motor control problems seen in humans and animals following a loss of dopamine cells. (C) 2009 Elsevier Ltd. All rights reserved.

There are two modes of control recognised in the cognitive psychological literature. Controlled processing is slow, requires serial attention to sub-tasks, and requires effortful memory retrieval and decision making. In contrast automatic control is less effortful, less prone to interference from simultaneous tasks, and is driven largely by the current stimulus. Neurobiological analogues of these are goal-directed and habit-based behaviour respectively. Here, we suggest how these control modes might be deployed in an engineering solution to Automatic Vehicle Control. We present pilot data on a first step towards instantiating automatised control in the architecture, and suggest a synergy between the engineering and biological investigation of this dual-process approach.

This paper describes emergent neurobiological characteristics of an intelligent multiple-controller that has been developed for controlling the throttle, brake and steering subsystems of a validated vehicle model. Simulation results demonstrate the effectiveness of the proposed approach. Importantly, the controller exhibits discrete behaviours, governed by its two component controllers. These controllers are selected according to task demands by a fuzzy-logic based supervisor. The system therefore displays `action selection' under central switched control, as has been proposed to take place in the vertebrate brain. In addition, the supervisor and modular controllers have analogues with the higher and lower levels of functionality associated with the strata in layered brain architectures. Several further similarities are identified between the biology and the vehicle controller. We conclude that advances in neuroscience and control theory have reached a critical mass which make it timely for a new rapprochement of these disciplines.

We present an application of a massively parallel processor array VLSI circuit to the implementation of neural networks in complex architectural arrangements. The work was motivated by existing biologically plausible models of a set of sub-cortical nuclei -the basal ganglia. The model includes 5 layers, each consisting of 16384 leaky integrator neurons, with inter-layer synaptic weights forming various one-to-one and diffuse connectivity patterns. The architecture of the SIMD processor array allows all the neurons per layer to be updated simultaneously. The performance of the processor array chip in simulating the model is compared with the original model being executed on a computer workstation. It is demonstrated that in this application the chip outperforms the workstation by five orders of magnitude in terms of computational performance and seven orders of magnitude in terms of energy efficiency, providing a high-speed, low-power, compact hardware platform for possible embedded robotic applications.

The implementation of a tactile sensory system for neuromorphic signal processing applications is presented. It has been developed by modelling the structure and behaviour of real rodent facial vibrissae. The primary afferents have been modelled using empirical data taken from electrophysiological measurements, and implemented in real time using hardware processors (DSP/FPGA). Pipelining techniques were employed to maximise the utility of the FPGA hardware. The system is to be integrated into a more complete whisker sensory model, including neural structures within the central nervous system, which can be used to orientate a mobile robot.

Both a FPGA implementation of a rodent?s tactile sensory system and a neural FPGA based hardware processor, mimicking the brainstem behaviour are presented. They have principally been designed using biological considerations. The two systems are being ported on a single FPGA platform and will ultimately be embedded on a mobile robot, which will operate in real world environments for object recognition and surface textural discrimination purposes.

The implementation of a large scale, leaky-integrate-and-fire neural network processor using the Xilinx Virtex-II family of field programmable gate array (FPGA) is presented. The processor has been designed to model biologically plausible networks of spiking neurons in real-time to assist with the control of a mobile robot. The real-time constraint has led to a re-evaluation of some of the established architectural and algorithmic features of previous spiking neural network based hardware. The design was coded and simulated using Handel-C hardware description language (HDL) and the DK3 design suite from Celoxica. The processor has been physically implemented and tested on a RC200 development board, also from Celoxica.

A real-time, large scale, leaky-integrate-and-fire neural network processor realized using FPGA is presented. This has been designed, as part of a collaborative project, to investigate and implement biologically plausible models of the rodent vibrissae based somatosensory system to control a robot. An emphasis has been made on hard real-time performance of the processor, as it is to be used as part of a feedback control system. This has led to a revision of some of the established modelling protocols used in other hardware spiking neural network processors. The underlying neuron model has the ability to model synaptic noise and inter-neural propagation delays to provide a greater degree of biological plausibility. The processor has been demonstrated modelling real neural circuitry in real-time, independent of the underlying neural network activity.

International audience; The basal ganglia system has been proposed as a possible neural substrate for action selection in the vertebrate brain. We describe a robotic implementation of a model of the basal ganglia and demonstrate the capacity of this system to generate adaptive switching between several acts when embedded in a robot that has to ‘survive' in a laboratory environment. A comparison between this brain-inspired selection mechanism and classical `winner-takes-all\' selection highlights some adaptive properties specific to the model, such as avoidance of dithering and energy-saving. These properties derive, in part, from the capacity of simulated basal ganglia-thalamo-cortical loops to generate appropriate ‘behavioural persistence'.

A self-organizing neural network is presented which uses, as input, Fourier information about the image of moving objects to represent the magnitude and orientation of their velocity vectors topologically over the neural layer. The input space consists of the outputs of spatio-temporal frequency and orientation tuned filters. The motion that is encoded is that in the plane perpendicular to the line of sight. The neural model of velocity encoding incorporates biologically plausible elements into a computational theory based on Fourier analysis of the image. At the systems level, the model may be represented as a two-stage process. The first stage is composed of input cell filters which extract information in the Fourier domain, and the second stage is the velocity encoding net itself. >

Under a wide range of conditions, stimuli composed of two superimposed grating components with unequal rotation velocities (differing in sign and/or magnitude) gave a striking percept of a single, coherent, nonrigidly deforming plaid surface. Conversely, component angular velocities of the same sign and magnitude yielded rigidly rotating plaids. Rigidity and motion coherence were shown to be independent percepts, and coherent plaids were categorised unambiguously as rigid or nonrigid. Coherence and motion transparency were found to depend upon the relative orientation of components, and polar plots showed two lobes of high coherence for narrow intercomponent angles. There was a slight tendency for plaids with unequal component rotations to appear less coherent, but this was nonsignificant, once the effect of intercomponent angle was taken into account. Changes in the relative spatial frequency of components, relative contrast of components, and repeated presentation produced equivalent effects on coherence for rigid and nonrigid types of plaid motion. Manipulation of the terminators in the display by making the aperture diameters for the two component gratings unequal reduced coherence and increased transparency. The effect was the same for rigid and nonrigid plaids. Coherence in rigid and nonrigid plaids thus depends primarily on local processes and there is no strong interaction between rigidity and coherence.