Michael Recce

Memory for places: A navigational model in support of Marr's theory of hippocampal function

In this report we describe a model that applies Marrs theory of hippocampal function to the problem of map‐based navigation. Like many others we attribute a spatial memory function to the hippocampus, but we suggest that the additional functional components required for map‐based navigation are located elsewhere in the brain. One of the key functional components in this model is an egocentric map of space, located in the neocortex, that is continuously updated using ideothetic (self‐motion) information. The hippocampus stores snapshots of this egocentric map. The modeled activity pattern of head direction cells is used to set the best egocentric map rotation to match the snapshots stored in the hippocampus, resulting in place cells with a nondirectional firing pattern. We describe an evaluation of this model using a mobile robot and demonstrate that with this model the robot can recognize an environment and find a hidden goal. This model is discussed in the context of prior experiments that were designed to discover the map‐based spatial processing of animals. We also predict the results of further experiments.

High speed vision-based quality grading of oranges

We describe a novel system for grading oranges into three quality bands, according to their surface characteristics. This processing operation is currently the only non-automated step in citrus packing houses. The system must handle fruit with a wide range of size (55-100 mm), shape (spherical to highly eccentric), surface coloration and defect markings. Furthermore, the point of stem attachment (the calyx) must be recognised in order to distinguish it from defects. A neural network classifier on rotation invariant transformations (Zernike moments) is used to recognise radial colour variation, that is shown to be a reliable signature of the stem region. This application requires both high throughput (5-10 oranges per second) and complex pattern recognition. Three separate algorithmic components are used to achieve this, together with state-of-the-art processing hardware and novel mechanical design. The grading is achieved by simultaneously imaging the fruit from six orthogonal directions as they are propelled through an inspection chamber. In the first stage processing colour histograms from each view of an orange are analysed using a neural network based classifier. Views that may contain defects are further analysed in the second stage using five independent masks and a neural network classifier. The computationally expensive stem detection process is then applied to a small fraction of the collected images. The succession of oranges constitute a pipeline, and, time saved in the processing of defect free oranges is used to provide additional time for other oranges. Initial results are presented from a performance analysis of this system.

Experimental modelling of time-of-flight sonar

This paper describes empirical models of sonar time-of-flight, derived from a large quantity of data collected with a single Polaroid ultrasonic rangefinder. We give a closed form model for the mean and standard deviation of range readings from a rough brick wall as a function of normal distance and beam incidence angle. For specular walls, we give the probability of direct (as opposed to multiple specular) return, and the mean and standard deviation of direct return range readings, as a function of incidence angle in tabular form. The models presented in this paper extend and correct those currently used in many robot sonar systems, particularly in the case of oblique incidence angles.

Vision and neural control for an orange harvesting robot

We describe the system control architecture of a large orange harvesting robot. This robot has two independent electrically driven telescopic arms mounted on a common platform which is itself held by a large hydraulic arm. This arm, in turn, is mounted on a tracked vehicle. The telescopic arms have cameras within the end-effecters, which are used to detect and measure the position and distance of the fruit within the canopy of a tree. Most of the development and control software was implemented using the matrix-based Virtual Machine Language (VML). This language was designed to implement neural networks, and has been extended and enhanced for robotic applications and the particular low level control requirements of the hardware. The device drivers provide the interface to frame grabbers, motor drivers, digital interface electronics, proximity detectors, and file handling. The same interface is used to implement interprocess communications with display and monitoring tools.

Toward a Mechanism for Navigation by the Rat Hippocampus

A simulation of the rat hippocampus as a mechanism for representing spatial information which is used to guide navigation is presented. A neuronal simulation of the firing patterns of four layers of cells: sensory, entorhinal, place and subicular cells, and a postulated set of goal cells form the basis of the model. Each cell type is characterised by the rat and timing of action potentials relative to the clock cycles provided by the hippocampal ? rhythm. Learning occurs in binary synapses, switched on by simultaneous pre- and post-synaptic activity. Activation spreads forward, through each layer in turn, to the goal cells which receive a reinforcement signal whenever a goal (e.g. food reward) is encountered. The ‘population vectors’ of sub-sets of goal cells code for the instantaneous direction of the rat from previously encountered rewards, allowing successful navigation in open fields.

Flexible Operating Environment for Matrix Based Neurocomputers

This paper describes the design and implementation of a development environment for matrix based neurocomputers. A new virtual machine language provides a wide range of matrix operations and device-related input/output communications. Virtual machines may be implemented entirely on conventional workstations or may use matrix-based neurocomputer hardware. To assist in algorithm development and debugging, the virtual machine is able to generate monitoring messages. A graphical interface is used to view the workings of one or more virtual machines. The user interface allows a range of display techniques to be associated with VML scalar and matrix variables. Virtual machines and monitoring processes run under the control of a central scheduler. All communications are implemented using a message based protocol. This environment is currently being used to develop a wide range of applications.

Neural model of a grid-based map for robot sonar

A functional similarity is described between cells of an occupancy grid for robot sonar, and integrate-and-fire neurons of an artificial neural net. Using this analogy, a new grid-based mapping system for robot sonar is described, which makes use of the neural concepts of receptive fields and recurrent connections. The performance of the new network is compared to that of a previous Bayesian grid-based mapping method, and a previous feature-based mapping method.

Interneuron plasticity in associative networks

Associative network models with binary synapses are widely studied as a biologically plausible memory mechanism. These models often include a single interneuron, used to set a global threshold for a network of sparsely interconnected principal cells, and the storage capacity improves with the use of a multi-step recall process (Gardner-Medwin, 1976). We demonstrate that the inclusion of non-saturating modifiable Hebbian synaptic weights in the projection from the interneuron to the principal cells drastically improves the performance of the network. These synaptic weights reduce the influence of the principal cells that are active in a disproportionate number of memory events.