Guido Bugmann

NEURAL NETWORK DESIGN FOR ENGINEERING APPLICATIONS

Abstract Computers are an integral part of day to day activities in engineering design and engineers have utilised various applications to assist them improve their design. Although computers are used to model a variety of engineering activities, currently the main focus of computer applications are areas with well defined rules. Activities related to the conceptual stage of the design process are generally untouched by computers. Artificial neural networks (ANN) have recently been widely used to model some of the human activities in many areas of science and engineering. Early applications of NN in civil engineering occurred the late eighties. One of the distinct characteristics of the ANN is its ability to learn from experience and examples and then to adapt with changing situations. Engineers often deal with incomplete and noisy data, which is one area where NN are most applicable. This is particularly the case at the conceptual stage of the design process. This paper presents practical guidelines for designing ANN for engineering applications. A brief introduction to NN is given; major aspects of three types of NN, multi-layer perceptron (MLP), radial basis network (RBF) and normalised RBF (NRBF) are discussed; new methods for selection and normalisation of training data are introduced and a practical example of a reinforced concrete slab design is presented.

Classification of Finger Movements for the Dexterous Hand Prosthesis Control With Surface Electromyography

A method for the classification of finger movements for dexterous control of prosthetic hands is proposed. Previous research was mainly devoted to identify hand movements as these actions generate strong electromyography (EMG) signals recorded from the forearm. In contrast, in this paper, we assess the use of multichannel surface electromyography (sEMG) to classify individual and combined finger movements for dexterous prosthetic control. sEMG channels were recorded from ten intact-limbed and six below-elbow amputee persons. Offline processing was used to evaluate the classification performance. The results show that high classification accuracies can be achieved with a processing chain consisting of time domain-autoregression feature extraction, orthogonal fuzzy neighborhood discriminant analysis for feature reduction, and linear discriminant analysis for classification. We show that finger and thumb movements can be decoded accurately with high accuracy with latencies as short as 200 ms. Thumb abduction was decoded successfully with high accuracy for six amputee persons for the first time. We also found that subsets of six EMG channels provide accuracy values similar to those computed with the full set of EMG channels (98% accuracy over ten intact-limbed subjects for the classification of 15 classes of different finger movements and 90% accuracy over six amputee persons for the classification of 12 classes of individual finger movements). These accuracy values are higher than previous studies, whereas we typically employed half the number of EMG channels per identified movement.

Mobile robot programming using natural language

Abstract How will naive users program domestic robots? This paper describes the design of a practical system that uses natural language to teach a vision-based robot how to navigate in a miniature town. To enable unconstrained speech the robot is provided with a set of primitive procedures derived from a corpus of route instructions. When the user refers to a route that is not known to the robot, the system will learn it by combining primitives as instructed by the user. This paper describes the components of the Instruction-Based Learning architecture and discusses issues of knowledge representation, the selection of primitives and the conversion of natural language into robot-understandable procedures.

Normalized Gaussian Radial Basis Function networks

Abstract The performances of normalised RBF (NRBF) nets and standard RBF nets are compared in simple classification and mapping problems. In normalized RBF networks, the traditional roles of weights and activities in the hidden layer are switched. Hidden nodes perform a function similar to a Voronoi tessellation of the input space, and the output weights become the networks output over the partition defined by the hidden nodes. Consequently, NRBF nets lose the localized characteristics of standard RBF nets and exhibit excellent generalization properties, to the extent that hidden nodes need to be recruited only for training data at the boundaries of class domains. Reflecting this, a new learning rule is proposed that greatly reduces the number of hidden nodes needed in classification tasks. As for mapping applications, it is shown that NRBF nets may outperform standard RBFs nets and exhibit more uniform errors. In both applications, the width of basis functions is uncritical, which makes NRBF nets easy to use.

Training personal robots using natural language instruction

As domestic robots become pervasive, uninitiated users will need a way to instruct them to adapt to their particular needs. The authors are designing a practical system that uses natural language to instruct a vision-based robot.

Role of temporal integration and fluctuation detection in the highly irregular firing of a leaky integrator neuron model with partial reset

Partial reset is a simple and powerful tool for controlling the irregularity of spike trains fired by a leaky integrator neuron model with random inputs. In particular, a single neuron model with a realistic membrane time constant of 10 ms can reproduce the highly irregular firing of cortical neurons reported by Softky and Koch (1993). In this article, the mechanisms by which partial reset affects the firing pattern are investigated. Itisshown theoretically that partial reset is equivalent to the use of a time-dependent threshold, similar to a technique proposed by Wilbur and Rinzel (1983) to produce high irregularity. This equivalent model allows establishing that temporal integration and fluctuation detection can coexist and cooperate to cause highly irregular firing. This study also reveals that reverse correlation curves cannot be used reliably to assess the causes of firing. For instance, they do not reveal temporal integration when it takes place. Further, the peak near time zero does not always indicate coincidence detection. An alternative qualitative method is proposed here for that later purpose. Finally, it is noted that as the reset becomes weaker, the firing pattern shows a progressive transition from regular firing, to random, to temporally clustered, and eventually to bursting firing. Concurrently the slope of the transfer function increases. Thus, simulations suggest a correlation between high gain and highly irregular firing.

Vision-based urban navigation procedures for verbally instructed robots

Abstract When humans explain a task to be executed by a robot they decompose it into chunks of actions. These form a chain of search-and-act sensory-motor loops that exit when a condition is met. In this paper we investigate the nature of these chunks in an urban visual navigation context, and propose a method for implementing the corresponding robot primitives such as “take the nth turn right/left”. These primitives make use of a “short-lived” internal map updated as the robot moves along. The recognition and localisation of intersections is done in the map using task-guided template matching. This approach takes advantage of the content of human instructions to save computation time and improve robustness.

Biologically plausible neural computation.

The function of a neuron can be described simultaneously at several levels of abstraction. For instance, a spike train represents the result of a computation done by a single neuron with its inputs, but it also represents the result of a complex function realized by the network in which the neuron is embedded. When models of large parts of the brain are considered, it may be desirable to use computational modules operating at a very abstract level. However, it is shown here that abstract neural functions depend on detailed features of the single neuron model used in the network reproducing the abstract function. Examples are given of the multiplicative function, motion detection, short-term memory and timing. All these operations rely on one or another feature of the extended Leaky Integrate-and-Fire neuron used in this paper, e.g. probabilistic synapses, post-synaptic currents modelled with alpha functions or partial reset of the somatic membrane. Consequently it is suggested that neural modelling at an abstract level does not obviate the need for a clear statement on the nature of the underlying model of biological neuron. In that sense, not many abstract functions are convincingly grounded, not even the standard formal neurons used in most artificial neural networks.

Improving the Performance Against Force Variation of EMG Controlled Multifunctional Upper-Limb Prostheses for Transradial Amputees

We investigate the problem of achieving robust control of hand prostheses by the electromyogram (EMG) of transradial amputees in the presence of variable force levels, as these variations can have a substantial impact on the robustness of the control of the prostheses. We also propose a novel set of features that aim at reducing the impact of force level variations on the prosthesis controlled by amputees. These features characterize the EMG activity by means of the orientation between a set of spectral moments descriptors extracted from the EMG signal and a nonlinearly mapped version of it. At the same time, our feature extraction method processes the EMG signals directly from the time-domain to reduce computational cost. The performance of the proposed features is tested on EMG data collected from nine transradial amputees performing six classes of movements each with three force levels. Our results indicate that the proposed features can achieve significant reductions in classification error rates in comparison to other well-known feature extraction methods, achieving improvements of ≈ 6% to 8% in the average classification performance across all subjects and force levels, when training with all forces.

A spiking neuron model: applications and learning

This paper presents a biologically inspired, hardware-realisable spiking neuron model, which we call 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. The TNLI incorporates temporal dynamics at the neuron level by modelling both the temporal summation of dendritic postsynaptic currents which have controlled delay and duration and the decay of the somatic potential due to its membrane leak. Moreover, the TNLI models the stochastic neurotransmitter release by real neuron synapses (with probabilistic RAMs at each input) and the firing times including the refractory period and action potential repolarisation. The temporal features of the TNLI make it suitable for use in dynamic time-dependent tasks like its application as a motion and velocity detector system presented in this paper. This is done by modelling the experimental velocity selectivity curve of the motion sensitive H1 neuron of the visual system of the fly. This application of the TNLI indicates its potential applications in artificial vision systems for robots. It is also demonstrated that Hebbian-based learning can be applied in the TNLI for postsynaptic delay training based on coincidence detection, in such a way that an arbitrary temporal pattern can be detected and recognised. The paper also demonstrates that the TNLI can be used to control the firing variability through inhibition; with 80% inhibition to concurrent excitation, firing at high rates is nearly consistent with a Poisson-type firing variability observed in cortical neurons. It is also shown with the TNLI, that the gain of the neuron (slope of its transfer function) can be controlled by the balance between inhibition and excitation, the gain being a decreasing function of the proportion of inhibitory inputs. Finally, in the case of perfect balance between inhibition and excitation, i.e. where the average input current is zero, the neuron can still fire as a result of membrane potential fluctuations. The firing rate is then determined by the average input firing rate. Overall this work illustrates how a hardware-realisable neuron model can capitalise on the unique computational capabilities of biological neurons.