We developed a portable biomimetic retina for blind subjects with retinal defects. 1-To simplify the visual prosthetic function, image segmentation of input patterns PI was provided by a set of line elements with specific lengths and orientations. 2-The segmented images were mapped by a filter module (FM: array of tunable spatio-temporal (ST) filters) onto a data stream as future stimulation signals for the human central visual system. 3-The foveal region of the central visual system was simulated by an inverter module (IM) to test the generation of a visual percept P2 of a given PI before the entire system will be applied to blind humans. 4-The parameter vector (PV) of FM could be modified interactively by the human user with evolutionary algorithms (EA) based on a perceptual comparison. 5-Two small displays for separate presentation of PI and the simulated percept P2 were integrated in a lightweight head mount and were combined with a 3-D acceleration sensor (AS) for head movement detection. 6-Subjects with normal vision were able to tune FM in a perception-based dialog exclusively by means of specific small head movements and to iteratively select the best three out of six possible percepts P2 until the output of IM, P2 became identical to a given PI.

This paper combines recent findings from neuroscience of the primate visual system, neural computation simulations, and dialog-based man-machine tuning experiments in order to define the properties that are essential for novel interactive image acquisition and display systems for human visual augmentation challenges: e.g. a-Interactive Imagery Analysis, b-Visual Prosthetics, and c-Generation of realistic, dynamically explorable Objects. Technical systems need for elicitation of visual percepts in humans a detailed consideration of sensory and motor parameters of the human visual system and also learning control capabilities to provide an optimization of visual percepts in a feedback loop between a technical system and the human.

In this paper, we present an optimization method for a learning algorithm for generation of tactile stimuli which are adapted by means of tactile perception of a human. Because of special requirements for tactile perception tuning the optimization of the proposed learning algorithm cannot be performed basing on gradient-descent or likelihood estimation methods. Therefore, an automatic tactile classification (ATC) is introduced for the optimization process. The results show that the ATC equals the tactile comparison of humans and that the learning algorithm is successfully optimized by means of the developed ATC.

In this paper we present an optimization method for a learning algorithm for tactile stimuli generation which are adapted by means of tactile perception of a human. Because of special requirements for a learning algorithm for tactile perception tuning the optimization cannot be performed based on gradient-descent or likelihood estimation methods. Therefore an automatic tactile classification (ATC) is introduced for the optimization process. The results show that the ATC equals the tactile comparison of humans and that the learning algorithm is successfully optimized by means of the ATC.

Summary form only given. From a systems theory and computational neuroscience perspective, the primate foveal visual system in the photopic range consists of a retina module as a large ensemble of spatio-temporal (ST) filters represented by the receptive field (RF) properties of mostly P- and Mganglion cells feeding into a corresponding central visual system module (VM). VM in turn elicits visual percepts P2 corresponding to optical input patterns P1. Human visual perception, which transcends neuroscience and biophysics, is considered here as the result of a sequence of two unidirectional mapping operations. The paper outlines a novel retina encoder (RE*) for mapping of optical patterns P1 onto vectors of ambiguous output signals; RE* serves both as retina module simulator and as neuroprosthetic retinal replacement. The paper also identifies essential requirements for the mapping of an ambiguous signal vector onto an unambiguous pattern. It also discusses perceptual consequences of a low-dimensional (e.g. 100) vector of multiple ganglion cell activity generated by RE* in blind subjects with an epiretinal, learning retina implant, vs. the high-dimensional vector of single ganglion cell activity generated by the human retina during normal vision.

Within This work a new data clustering algorithm is proposed based on classical clustering algorithms. Here k-means neurons are used as substitute for the original data points. These neurons are online adaptable extending the standard k-means clustering algorithm. They are equipped with perceptive fields to identify if a presented data pattern fits within its area it is responsible for. In order to find clusters within the input data an extension of the /spl isin/-nearest neighboring algorithm is used to find connected groups within the set of k-means neurons. Most of the information the clustering algorithm needs is taken directly from the input data. Thus only a small number of parameters have to be adjusted. The clustering abilities of the presented algorithm are shown using data sets from two different kinds of applications.

A retina encoder (RE) as part of a visual prosthesis (retina implant) for blind subjects with retinal degenerative disorders was implemented by an array of 256 tunable spatio-temporal filters (ST) to map visual patterns P1 onto encoded output patterns as functional approximation of retina information processing. These RE-output signals were fed into a visual system module (VM) to simulate the mapping (by the central visual system) of the retinal output onto visual percepts P2, which were visualized on a monitor. Alternative tuning strategies for roaming within the large spatio-temporal state space of RE were developed and tested. Initially, VM as neural network was trained to generate a P2 very similar to a selected set of P1s by feeding the corresponding RE-outputs into VM. RE tuning was tested by presenting a given small set of P1s to the serial coupling of RE and VM and by monitoring a modified Hamming distance between P2 and P1, while the ST-parameters were tuned. Human volunteers with normal vision were asked to approximate P2 to P1 directly, with random search, or by means of a dialog-based tuning with an evolutionary algorithm (EA). Typically, dialog-based EA tuning reached the optimal similarity between P2 and P1 (limited by the quality of VM) within less than 200 iterations. In contrast, manual tuning required considerably more iteration steps and converged to a smaller similarity between P2 and P1. Structure and function of our tunable RE offer not only an optimization of visual perception, but may also be important to serve as technical encryption unit (EU) in connection with the human central visual system as biological decryption unit (DU); an important feature to meet the authentication requirements of active medical devices.

We present a tactile intelligent sensory substitution system (TIS/sup 3/) as a novel tactile communication system with dialog-based tuning. TIS/sup 3/ consists of a tactile encoder (TE) which maps desired objects or patterns (P) onto spatio-temporal stimulation patterns (P') as a parallel stream of stimulation time courses, by means of a tactile stimulator array (TS) at selected skin location to elicit tactile perceptions (P*). The human subject evaluates and compares a percept P* with a given object P as association goal. A Learning Module (LM) for dialog based TE-tuning transforms these evaluations into TE-change signals. In a first step, the application of a dialog-based TE-tuning was successfully tested using a TS version with fifteen stimulators on the lower arm for regaining the corresponding stimulation pattern P' for a given tactile reference percept P*/sub ref/. Optimization of P* was achieved in less than 100 iteration steps by means of a micro evolutionary learning algorithm (MEA).

An image processing system for the real-time object detection and recognition was designed on the principles of active vision and sequential analysis. The real-world visual tasks can be solved due to predictive control of the vision sensor. The sequential analysis allows real-time implementation of the system on low cost DSP hardware. The system was implemented on the DSP TMS320C50 and requires 18-30 ms for the detection and recognition of the object.

We present a new design of hybrid force/position control NVFC capable of friction-compensation, which is a mandatory requirement in industrial applications for precise tracking of hard surfaces with desired contact force. The control architecture consists of an outer-loop velocity controller and an inner-loop adaptive hybrid force/position controller, which compensates the friction of the manipulators joints, significantly improving the force tracking performance during slow movements. The cascaded velocity controller ensures the precise approach during neural force/position control with a reduced bounce into the unknown surface, which is a demanding requirement for tasks like deburring or chamfering. The resulting performance of the force control system is illustrated in simulations, and is consistent with velocity/force control experiments on a 6-DOF industrial manipulator controlled with a PC-based real-time control system.

We present a novel control concept that solves a wide range of surface tracking tasks for manipulators with defined contact to (moving) rigid objects. The position based neural force control (NFC-P) consists of a hybrid force/position controller that accurately generates contact forces to objects with arbitrary flexibility and uncertain distance or shape. NFC-P performs force control by modifying the desired joint angle changes in force direction. These are fed into a computed torque controller, where the inverse dynamics of the manipulator is represented by neural networks. NFC-P includes a neural trajectory generating tool for smooth and kinematical valid contact trajectories with the possibility to adapt the trajectories to the unknown shape of the surface online. The kinematical mappings guarantee singularity robustness (SR) in the entire workspace. Results from real-time experiments are presented using a 6-DOF industrial manipulator as testbed.

We present a novel type of friction estimation applied to the field of sensorless force/position control. As part of a position based neural force control (NFC-P) the estimation friction and external force allows a force/position control without using a force sensor. NFC-P consists of a hybrid force/position controller that accurately generates contact forces to objects with arbitrary flexibility and uncertain distance or shape. NFC-P performs force control by modifying the desired joint angle changes in force direction before they are fed into a computed torque controller. The inverse dynamics of the manipulator is modeled in a computed torque controller. Kinematic mappings guarantee singularity robustness in the entire workspace. Results from real time experiments are presented with a 6-DOF industrial manipulator as a testbed.

The perception-based tuning procedure of a learning retina implant as a visual prosthesis for blind subjects with retinal degeneration was tested in subjects with normal vision. A simulated retina encoder (RE) with 256 tunable spatiotemporal receptive field (RF) filters was stimulated by spatiotemporal visual patterns. RE output impulse rate was fed into an inversion module (IM) to reproduce the corresponding input pattern in real time. Initially, IM was trained to generate an approximation of the inverse of the RE mapping for a given RE state S, so as to yield a close similarity between spatiotemporal RE input and IM output patterns. Subsequently, S was arbitrarily changed thus drastically increasing the difference between desired and perceived pattern. A learning algorithm based on evolution strategies was developed to adjust RE iteratively with human feedback.

A tuning method with reinforcement learning (RL) for the retina encoder (RE) of a retina implant (RI) as a visual prosthesis for blind subjects with retinal degenerations is proposed. The RE simulates retinal information processing in real time by means of spatio-temporal receptive field (RF) filters and generates electrical signals for stimulation of several hundreds of ganglion cells (GC) to regain a modest amount of vision. For each contacted GC, the RE has to be optimized with regard to the patient's perception. The patient's (for the present simulated) evaluative feedback is applied here in a dialog module as a reinforcement signal to train several RL agents in a neural network learning process.

We developed a novel concept of hybrid force/position control based on neural networks (NFC) to significantly expand the range of manipulator applications. NFC includes neural approaches for complex robotic mappings such as inverse dynamics and kinematics. The neural dynamics network, as an essential component of the computed torque controller, performs a fast and adaptive computation of the inverse manipulator model. The kinematic mappings are represented by a neural kinematics network (NKN). The features of NKN provide singularity robustness and the handling of constraints in joint space and Cartesian space. To guarantee a tender impact while establishing contact between manipulator and surface, a cascaded velocity controller is added to the NFC approach. Simulations for a 6-DOF industrial manipulator have proved that the NFC concept is capable to manage various demanding tasks such as screw removal and surface tracking with high accuracy.

We developed a generally applicable concept for flexible path planning and representation in high-dimensional configuration spaces. Therefore, an AI-algorithm for fast preprocessing and a neural network were combined. Specifically, the standard A*-method was developed into an advanced A*-method (AA*-method) by creating an additional class of FREE-cells to enlarge the computed surroundings of the detected optimal path, and by constituting expansion matrices to enable flexible modeling of different cell extents and configuration spaces. Furthermore, a neural RBF-network was modified by adding an activity peak generating neuron guaranteeing updates in real-time (less than 1 ms). The output of the AA*-method, a set of classified cells, was used to train the modified RBF-network. The capabilities of this novel hybrid path planning system are demonstrated for various complex 3D- and 6D- path planning tasks.

Coherent-type artificial neural networks whose behavior is controlled by carrier-frequency modulation are proposed. The network learns teacher signals associated with an information-carrier frequency as a network parameter. The total network system forms a self-homodyne circuit. The learning process is realized by adjusting delay time and conductance of neural connections. Experiments demonstrate that the network behavior is successfully controlled by the carrier-frequency modulation. This result will be applicable not only to signal processing but also to frequency-multiplexed optical neural computing and quantum neural devices such as carrier-energy-controlled neurons in the future.

A completely parallel, asynchronous, pulse-processing neural net has been described. This hardware simulator is fabricated with discrete electronic elements. A variety of basic biological parameters are efficiently incorporated in the hardware. It is noteworthy that the concept uses neural activity to modify synaptic weights and time delays. It follows that learning rules are embedded in the net topology

The authors present a network structure consisting of two parts with different topologies, the time-pattern generator (TPG) and the activity memory (AM). It is designed to generate an arbitrary temporal sequence as necessary for controlling actuators with artificial neural nets. The TPG represents the flow of time as the spatial location of a traveling activity peak (AP) initiated at the TPG input neurons by higher neural centers. The AM network modulates the output activity according to the transfer function of the AM neurons and the time-dependent sum of TPG outputs weighted with the synaptic strength of the connections between the TPG and the AM. The synaptic weights of the AM carry the information necessary for the generation of temporally varying amplitudes. As an example for the suggested network topology, the authors present a straightforward realization

A four-joint machine with possible movements in a horizontal, planar area is considered. A neural network mechanism to solve the problem of inverse kinematics for this system is constructed. The mechanism is based on the idea that the joint which is most suitable for generating a given trajectory element should contribute most to the actual movement. The possible contributions of the joints are linearized and scaled by a gain function due to the geometry of the physical four-joint machine. A hierarchical sequence determines an order and therefore the directions of function as flexor or extensor for the four actuators for each step along the trajectory. The hierarchical sequence specifies that the vectors of the contribution of the first two selected actuators are added up, whereas the remaining two joints are used to compensate the vector components perpendicular to the vector of the currently processed trajectory element. This is realized by deciding about the contribution of the various joints in a geometrical manner. Since this local mechanism provides a satisfactory solution only for some parts of the working area, a small amount of knowledge about the history of the join movement has to be incorporated to generate an overall good solution

A framework for two oculomotor subsystems, an electromechanical pursuit system and the saccadic system, is proposed on the basis of neurophysiological findings in alert monkeys. This model is arranged into a sequence of spatiotemporal translation, motor program generation, and neural integration. For the purpose of foveal pursuit movements, an eye velocity signal must be internally generated and updated according to the current position error on the retina. This design probably involves a recently discovered class of pursuit neurons. Saccadic eye movements are probably controlled by burst neurons. Based on striking similarities between the neural activity of these recently analyzed burst neurons and Renshaw cells, a detailed quantitative model for the generation of preprogrammed saccades, the Renshaw model, is proposed. Open and controversial questions in the literature concerning various strategies being used by the primate oculomotor system are mentioned.