No description available.

No description available.

The attempt to evolve complete embodied and situated artificial creatures in which both morphological and control characteristics are adapted during the evolutionary process has been and still represents a long term goal key for the artificial life and the evolutionary robotics community. Loosely inspired by ancient biological organisms which are not provided with a central nervous system and by simple organisms such as stick insects, this thesis proposes a new genotype encoding which allows development and evolution of mor- phology and neural controller in artificial agents provided with a distributed neural network. In order to understand if this kind of network is appropriate for the evolution of non trivial behaviours in artificial agents, two experiments (description and results will be shown in chapter 3) in which evolution was applied only to the controller’s parameters were performed. The results obtained in the first experiment demonstrated how distributed neural networks can achieve a good level of organization by synchronizing the output of oscillatory elements exploiting acceleration/deceleration mechanisms based on local interactions. In the second experiment few variants on the topology of neural architecture were introduced. Results showed how this new control system was able to coordinate the legs of a simulated hexapod robot on two different gaits on the basis of the external circumstances. After this preliminary and successful investigation, a new genotype encoding able to develop and evolve artificial agents with no fixed morphology and with a distributed neural controller was proposed. A second set of experiments was thus performed and the results obtained confirmed both the effectiveness of genotype encoding and the ability of distributed neural network to perform the given task. The results have also shown the strength of genotype both in generating a wide range of different morphological structures and in favouring a direct co-adaptation between neural controller and morphology during the evolutionary process. Furthermore the simplicity of the proposed model has showed the effective role of specific elements in evolutionary experiments. In particular it has demonstrated the importance of the environment and its complexity in evolving non-trivial behaviours and also how adding an independent component to the fitness function could help the evolutionary process exploring a larger space solutions avoiding a premature convergence towards suboptimal solutions.

No description available.

Short-term synaptic depression is the phenomena where repeated stimulation leads to a decreased transmission efficacy. In this thesis, the impact of synaptic depression on the responses and dynamics of network models of visual processing is investigated, and the coding implications are examined. I find that synaptic depression can fundamentally change the operation of previously well - understood networks, and explain temporal nonlinearities present in neural responses to multiple stimuli. Furthermore, I show, more generally, how nonlinear interactions can be beneficial with respect to neural coding. I begin chapter 1 with a short introduction. In chapter 2 of this thesis, the behaviour of a ring attractor network is examined when its recurrent connections are subject to short term synaptic depression. I find that, in the presence of a uniform background current, the activity of the network settles to one of three states: a stationary attractor state, a uniform state or a rotating attractor state. I show that the rotation speed can be adjusted over a large range by changing the background current, opening the possibility to use the network as a variable frequency oscillator or pattern generator, and use mathematical analysis to determine an approximate maximum rotation speed. Using simulations, I then extend the network into two - dimensions, and find a rich range of possible behaviours. Processing in the visual cortex can be non - linear: the response to two objects or other visual stimuli presented simultaneously is often less than the sum of the responses to the individual objects. A maximum function has in some cases been proposed to describe these competitive interactions. More recent data has emphasised that such interactions have temporal aspects as well, namely that the response to an initially presented stimulus can suppress the response to a stimulus presented subsequently, especially if the first stimulus is presented at high contrast. Chapter 3 of this thesis will present a simple neuronal network featuring synaptic depression which can account for much of the temporal aspects of this behaviour, whilst remaining consistent with older data and models. Furthermore, it will show how this model leads to several strong predictions regarding the processing of low contrast stimuli sequences, as well suggesting a link between response latency and suppression strength. The response of the model to a structured sequence of input stimuli also appears to anticipate future stimuli, and we predict that the magnitude of this stimulus anticipation will decrease as contrast is decreased. Following on from investigating the temporal aspects of responses to stimuli pairs, in chapter 4 this thesis examines an abstract model of how coding is impacted by non - linear interactions, for both structured and unstructured stimuli spaces. I find that non- linear methods of responding to pairs of stimuli presented simultaneously can have a beneficial effect on coding capacity, with linearly combined responses generally leading to the highest decoding errors rates. This thesis goes on to examine the interplay between this models noise assumptions and the decoding performance, and finds that many of the assumptions made can be weakened without changing, qualitatively, these findings. In chapter 5, this thesis examines layered networks of noisy spiking neurons with recurrent connectivity and featuring depressing synapses. The contrast dependent latency and spike count statistics of the model are analysed and are found to be strongly dependent on the parameters of the noise. The tuning of parameters for models containing noisy IF neurons is discussed, and an information theoretic approach to tuning is outlined which successfully reproduces earlier work in which noise was tuned to linearise the response of a spiking network. The approach is applied to maximise the ability of the network to filter rapid noise transients at low contrast. I finish the thesis with a short concluding chapter.

This thesis presents a Multiple Self-Organised Spiking Neural Networks (MSOSNN). The aim of this architecture is to achieve a more biologically plausible artificial neural network. Spiking neurons with delays are proposed to encode the information and perform computations. The proposed method is further implemented to enable unsupervised competitive and self-organising learning. The method is evaluated by application to real world datasets. Computer simulation results show that the proposed method is able to function similarly to conventional neural networks i.e. the Kohonen Self-Organising Maps. The SOSNN are further combined to form multiple networks of the Self-Organised Spiking Neural Networks. This network architecture is structured into <i>n</i> component modules with each module providing a solution to the sub-task and then combined with other modules to solve the main task. The training is made in such a way that a module becomes a winner at each step of the learning phase. The evaluation using different data sets as well as comparing the network to a single unity network showed that the proposed architecture is very useful for high dimensional input vectors. The Multiple SOSNN architecture thus provides a guideline for a complex large-scale network solution.

This thesis presents a way to generate a Digital Terrain Model (dtm) from a Digital Surface Model (dsm) and multi spectral images (including the Near Infrared (nir) color band). An Artificial Neural Network (ann) is used to pre-classify the dsm and multi spectral images. This in turn is used to filter the dsm to a dtm. The use of an ann as a classifier provided good results. Additionally, the addition of the nir color band resulted in an improvement of the accuracy of the classifier. Using the classifier, a dtm was easily extracted without removing natural edges or height variations in the forests and cities. These challenges are handled with great satisfaction as compared to earlier methods.

A key component of autonomous outdoor navigation in unstructured environments is the classification of terrain. Recent development in the area of machine learning show promising results in the task of scene segmentation but are limited by the labels used during their supervised training. In this work, we present and evaluate a flexible strategy for terrain classification based on three components: A deep convolutional neural network trained on colour, depth and infrared data which provides feature vectors for image segmentation, a set of exchangeable segmentation engines that operate in this feature space and a novel, air pressure based actuator responsible for distinguishing rigid obstacles from those that only appear as such. Through the use of unsupervised learning we eliminate the need for labeled training data and allow our system to adapt to previously unseen terrain classes. We evaluate the performance of this classification scheme on a mobile robot platform in an environment containing vegetation and trees with a Kinect v2 sensor as low-cost depth camera. Our experiments show that the features generated by our neural network are currently not competitive with state of the art implementations and that our system is not yet ready for real world applications.

The work described in this thesis is based on the results of a clinical trial conducted by the research team at the Medical Informatics Unit of the University of Cambridge, which show that the Ballistocardiogram (BCG) has prognostic value in detecting impaired left ventricular function before it becomes clinically overt as myocardial infarction leading to sudden death. The objective of this study is to develop and demonstrate a framework for realising an on-line BCG signal classification model in a portable device that would have the potential to find pathological signs as early as possible for home health care. Two new on-line automatic BeG classification models for time domain BeG classification are proposed. Both systems are based on a two stage process: input feature extraction followed by a neural classifier. One system uses a principal component analysis neural network, and the other a discrete wavelet transform, to reduce the input dimensionality. Results of the classification, dimensionality reduction, and comparison are presented. It is indicated that the combined wavelet transform and MLP system has a more reliable performance than the combined neural networks system, in situations where the data available to determine the network parameters is limited. Moreover, the wavelet transfonn requires no prior knowledge of the statistical distribution of data samples and the computation complexity and training time are reduced. Overall, a methodology for realising an automatic BeG classification system for a portable instrument is presented. A fully paralJel neural network design for a low cost platform using field programmable gate arrays (Xilinx's XC4000 series) is explored. This addresses the potential speed requirements in the biomedical signal processing field. It also demonstrates a flexible hardware design approach so that an instrument's parameters can be updated as data expands with time. To reduce the hardware design complexity and to increase the system performance, a hybrid learning algorithm using random optimisation and the backpropagation rule is developed to achieve an efficient weight update mechanism in low weight precision learning. The simulation results show that the hybrid learning algorithm is effective in solving the network paralysis problem and the convergence is much faster than by the standard backpropagation rule. The hidden and output layer nodes have been mapped on Xilinx FPGAs with automatic placement and routing tools. The static time analysis results suggests that the proposed network implementation could generate 2.7 billion connections per second performance.

Conventional feed forward Neural Networks have used the sum-of-squares cost function for training. A new cost function is presented here with a description length interpretation based on Rissanen's Minimum Description Length principle. It is a heuristic that has a rough interpretation as the number of data points fit by the model. Not concerned with finding optimal descriptions, the cost function prefers to form minimum descriptions in a naive way for computational convenience. The cost function is called the Naive Description Length cost function. Finding minimum description models will be shown to be closely related to the identification of clusters in the data. As a consequence the minimum of this cost function approximates the most probable mode of the data rather than the sum-of-squares cost function that approximates the mean. The new cost function is shown to provide information about the structure of the data. This is done by inspecting the dependence of the error to the amount of regularisation. This structure provides a method of selecting regularisation parameters as an alternative or supplement to Bayesian methods. The new cost function is tested on a number of multi-valued problems such as a simple inverse kinematics problem. It is also tested on a number of classification and regression problems. The mode-seeking property of this cost function is shown to improve prediction in time series problems. Description length principles are used in a similar fashion to derive a regulariser to control network complexity.