Olfactory Feature Maps from an Electronic Nose

The human brain contains approximately 1all nerve cells called "neurones". The main elements of each neurone are understood to be a cell body, an axon and a considerable number of dendrites. The cell body receives electrical signals for processing via its dendrites. The output signal is then transmitted along the axon which acts as a transmission line. This signal is finally passed across the synaptic links to the dendrites of other nerve cells. In this way electrical signals, and thus information, can pass from one neurone to a neighbouring one. The human brain has a complex and massive set (ca. 10,000) of interconnections between neurones. A neurone may also have a signal feedback system, with its axon connected to one or more of its own dendrites. Thus the output from a neurone can depend on a function of the inputs and/or the state of other neurones in the vicinity. These properties, and others, gives the human brain its highly parallel and sophisticated processing characteristics. The design of Artificial Neural Networks (ANNs) which mimic the biological nervous system has been inspired largely by the study of the human brain. The main reasons for building brain-like machines include the fact that they are highly parallel (this property should offer a high speed and efficiency of information processing at a relatively low clock speed); they learn by themselves; and information is stored in a distributed manner so there is greater fault-tolerance than conventional sequential microprocessor with discrete localised memory. Over the past ten years work has been carried out at Warwick University I to develop an instrument that mimics the human sense of smell. Our olfactory sense is the primary sense that we use to determine the flavour of foodstuffs or other substances. The olfactory system consists of a large number of olfactory receptor cells located in the olfactory epithelium which is high up in the nasal cavity. These cells are connected to other mitral cells via a complex interconnection layer of glomeruli nodes. Finally, the mitral cells feed synaptically into the main part of the brain for processing. The biological system has many features that we would like to incorporate in our electronic instrument. Past efforts have centred upon the use of an array of chemical solidstate gas sensors coupled to a microcomputer to obtain characteristic "fingerprints" of the associated sensor response [1]. Over recent years a considerable number of ANN paradigms have been developed including the popular MultiLayer Perceptrons (MLPs). The backpropagation MLP has been used as a predictive classifier of the output from our electronic nose with some success [2]. However, the backpropagation MLP requires, unlike the biological system, the supervised training of the neural network. On the other hand, the Kohonen Feature Map (KFM) is a paradigm which is said to have biological plausibility because it emulates the part of the brain which displays its associative characteristics [3,4], and this is the focus of the work presented here. A comprehensive discussion of an unsupervised version of KFM has been presented in a series of papers elsewhere [5-7], which describe the benefits of using KFM in the context of modelling time-dependent and drift mechanisms. These include some inherent features of KFM, such as a reduction in dimensionality (cf. principal component analysis) and invariance to drift and transitory noise, which are attributed to the convergent nature of a feature map. In this paper we not only consider unsupervised but also supervised versions of KFM. There are many ways of representing models of the artificial neurone. Two of the most popular are shown in Figures 1 and 2 [8]. Figure 1 shows the weighted summer with a non-linear activation functionF (e.g. sigmoid) and a threshold level q, where