Qingxiang Wu

A novel approach for the implementation of large scale spiking neural networks on FPGA hardware

This paper presents a strategy for the implementation of large scale spiking neural network topologies on FPGA devices based on the I&F conductance model. Analysis of the logic requirements demonstrate that large scale implementations are not viable if a fully parallel implementation strategy is utilised. Thus the paper presents an alternative approach where a trade off in terms of speed/area is made and time multiplexing of the neuron model implemented on the FPGA is used to generate large network topologies. FPGA implementation results demonstrate a performance increase over a PC based simulation.

Edge Detection Based on Spiking Neural Network Model

Inspired by the behaviour of biological receptive fields and the human visual system, a network model based on spiking neurons is proposed to detect edges in a visual image. The structure and the properties of the network are detailed in this paper. Simulation results show that the network based on spiking neurons is able to perform edge detection within a time interval of 100 ms. This processing time is consistent with the human visual system. A firing rate map recorded in the simulation is comparable to Sobel and Canny edge graphics. In addition, the network can separate different edges using synapse plasticity, and the network provides an attention mechanism in which edges in an attention area can be enhanced.

Learning under weight constraints in networks of temporal encoding spiking neurons

Limits on synaptic efficiency are characteristic of biological neural networks. In this paper, weight limitation constraints are applied to the spike time error-backpropagation (SpikeProp) algorithm for temporally encoded networks of spiking neurons. A novel solution to the problem raised by non-firing neurons is presented which makes the learning algorithm converge reliably and efficiently. In addition a square cosine encoder is applied to the input neurons to reduce the number of input neurons required. The approach is demonstrated by application to the classical XOR-problem analysis, a function approximation experiment and benchmark data sets. Using input delay neurons and relative timing, the algorithm is also applied to solve a time series prediction problem. The experimental results show that the new approach produces comparable accuracy in classification with the original approach while utilising a smaller spiking neural network.

Multiknowledge for decision making

The representation of knowledge has an important effect on automated decision-making. In this paper, vector spaces are used to describe a condition space and a decision space, and knowledge is represented by a mapping from the condition space to the decision space. Many such mappings can be obtained from a training set. A set of mappings, which are created from multiple reducts in the training set, is defined as multiknowledge. In order to get a good reduct and find multiple reducts, the WADF (worst-attribute-drop-first) algorithm is developed through analysis of the properties of decision systems using rough set theory. An approach that combines multiknowledge and the naïve Bayes classifier is applied to make decisions for unseen instances or for instances with missing attribute values. Benchmark data sets from the UCI Machine Learning Repository are used to test the algorithms. The experimental results are encouraging; the prediction accuracy for unseen instances by using the algorithms is higher than by using other approaches based on a single body of knowledge.

Multi-knowledge extraction and application

Rough set theory provides approaches to the finding a reduct (informally, an identifying set of attributes) from a decision system or a training set. In this paper, an algorithm for finding multiple reducts is developed. The algorithm has been used to find the multi-reducts in data sets from UCI Machine Learning Repository. The experiments show that many databases in the real world have multiple reducts. Using the multi-reducts, multiknowledge is defined and an approach for extraction is presented. It is shown that a robot with multi-knowledge has the ability to identify a changing environment. Multiknowledge can be applied in many application areas in machine learning or data mining domain.

A visual attention model based on hierarchical spiking neural networks

Based on the information processing functionalities of spiking neurons, hierarchical spiking neural networks are proposed to simulate visual attention. Using spiking neural networks inspired by the visual system, an image can be decomposed into multiple visual image components. Based on specific visual image components and image features, a visual attention system is proposed to extract attention areas according to top–down volition-controlled signals. The hierarchical spiking neural networks are constructed with a conductance-based integrate-and-fire neuron model and a set of specific receptive fields in different levels. The simulation algorithm and properties of the networks are detailed in this paper. Simulation results show that the attention system is able to perform visual attention of objects based on specific image components or features, and a demonstration shows how the attention system can detect a house in a visual image. Using the proposed saliency index, attention areas of interest can be extracted from spike rate maps of multiple visual pathways, such as ON/OFF colour pathways. According to this visual attention principle, the visual image processing system can quickly focus on specific areas while ignoring other areas.

Remembering Key Features of Visual Images Based on Spike Timing Dependent Plasticity of Spiking Neurons

The brain has the powerful capability of remembering key features of images. Based on the principle of spike timing dependent plasticity of spiking neurons and the ON/OFF pathways in the visual system, a spiking neural network is proposed to remember key features of visual images. The simulation results show that the network is capable of remembering key features according to a learning rule based on spike timing dependent plasticity. The principle of the network can be used to explain how a spiking neuron-based system can store the key features of visual images. Furthermore, the network can be applied to spiking neuron based artificial intelligent systems to support the processing visual images.

2D co-ordinate transformation based on a spike timing-dependent plasticity learning mechanism

In order to plan accurate motor actions, the brain needs to build an integrated spatial representation associated with visual stimuli and haptic stimuli. Since visual stimuli are represented in retina-centered co-ordinates and haptic stimuli are represented in body-centered co-ordinates, co-ordinate transformations must occur between the retina-centered co-ordinates and body-centered co-ordinates. A spiking neural network (SNN) model, which is trained with spike-timing-dependent-plasticity (STDP), is proposed to perform a 2D co-ordinate transformation of the polar representation of an arm position to a Cartesian representation, to create a virtual image map of a haptic input. Through the visual pathway, a position signal corresponding to the haptic input is used to train the SNN with STDP synapses such that after learning the SNN can perform the co-ordinate transformation to generate a representation of the haptic input with the same co-ordinates as a visual image. The model can be applied to explain co-ordinate transformation in spiking neuron based systems. The principle can be used in artificial intelligent systems to process complex co-ordinate transformations represented by biological stimuli.

Evolutionary Design Of Spiking Neural Networks

Unlike traditional artificial neural networks (ANNs), which use a high abstraction of real neurons, spiking neural networks (SNNs) offer a biologically plausible model of realistic neurons. They differ from classical artificial neural networks in that SNNs handle and communicate information by means of timing of individual pulses, an important feature of neuronal systems being ignored by models based on rate coding scheme. However, in order to make the most of these realistic neuronal models, good training algorithms are required. Most existing learning paradigms tune the synaptic weights in an unsupervised way using an adaptation of the famous Hebbian learning rule, which is based on the correlation between the pre- and post-synaptic neurons activity. Nonetheless, supervised learning is more appropriate when prior knowledge about the outcome of the network is available. In this paper, a new approach for supervised training is presented with a biologically plausible architecture. An adapted evolutionary strategy (ES) is used for adjusting the synaptic strengths and delays, which underlie the learning and memory processes in the nervous system. The algorithm is applied to complex non-linearly separable problems, and the results show that the network is able to perform learning successfully by means of temporal encoding of presented patterns.

A Distribution-Index-Based Discretizer for Decision-Making with Symbolic AI Approaches

When symbolic AI approaches are applied to handle continuous valued attributes, there is a requirement to transform the continuous attribute values to symbolic data. In this paper, a novel distribution-index-based discretizer is proposed for such a transformation. Based on definitions of dichotomic entropy and a compound distributional index, a simple criterion is applied to discretize continuous attributes adaptively. The dichotomic entropy indicates the homogeneity degree of the decision value distribution, and is applied to determine the best splitting point. The compound distributional index combines both the homogeneity degrees of attribute value distributions and the decision value distribution, and is applied to determine which interval should be split further; thus, a potentially improved solution of the discretization problem can be found efficiently. Based on multiple reducts in rough set theory, a multiknowledge approach can attain high decision accuracy for information systems with a large number of attributes and missing values. In this paper, our discretizer is combined with the multiknowledge approach to further improve decision accuracy for information systems with continuous attributes. Experimental results on benchmark data sets show that the new discretizer can improve not only the multiknowledge approach, but also the naive Bayes classifier and the C5.0 tree