Sensor-based machine olfaction with neuromorphic models of the olfactory system

Abstract

Electronic noses combine an array of cross-selective gas sensors with a pattern recognition engine to identify odors. Pattern recognition of multivariate gas sensor response is usually performed using existing statistical and chemometric techniques. An alternative solution involves developing novel algorithms inspired by information processing in the biological olfactory system. The objective of this dissertation is to develop a neuromorphic architecture for pattern recognition for a chemosensor array inspired by key signal processing mechanisms in the olfactory system. Our approach can be summarized as follows. First, a high-dimensional odor signal is generated from a chemical sensor array. Three approaches have been proposed to generate this combinatorial and high dimensional odor signal: temperature-modulation of a metal-oxide chemoresistor, a large population of optical microbead sensors, and infrared spectroscopy. The resulting high-dimensional odor signals are subject to dimensionality reduction using a self-organizing model of chemotopic convergence. This convergence transforms the initial combinatorial high-dimensional code into an organized spatial pattern (i.e., an odor image), which decouples odor identity from intensity. Two lateral inhibitory circuits subsequently process the highly overlapping odor images obtained after convergence. The first shunting lateral inhibition circuits perform gain control enabling identification of the odorant across a wide range of concentration. This shunting lateral inhibition is followed by an additive lateral inhibition circuit with center-surround connections. These circuits improve contrast between odor images leading to more sparse and orthogonal patterns than the one available at the input. The sharpened odor image is stored in a neurodynamic model of a cortex. Finally, anti-Hebbian/ Hebbian inhibitory feedback from the cortical circuits to the contrast enhancement circuits performs mixture segmentation and weaker odor/background suppression, respectively. We validate the models using experimental datasets and show our results are consistent with recent neurobiological findings.