Performance of a Computational Model of the Mammalian Olfactory System

Sensory inputs allow the brain to construct a detailed multimodal map of the surrounding world. Smell is a carrier of critical information about food, predators, and social status, among others. Thus, a wide range of mammalian or insect behaviors rely on the recognition of relevant odor stimuli. In the olfactory system, it involves the identification of a composition of molecules in rich odorant mixtures. Considering ecological aspects of this fundamental sensory task, in particular the complexity of naturally occurring odors blended in the chemically noisy environments and variability in olfactory stimulation conditions, manifested in the flux of perceived quality and intensity, the objective to maintain stable object perception is challenging. In order to construct the percept, the system must detect a chemical stimulus, extract its relevant features from fluctuating odiferous backgrounds, create their robust representations, and match them with odor patterns experienced earlier and stored in memory (Cleland and Linster 2005). The fundamental computational challenges in robust odor recognition beyond simple discrimination, which biological olfaction successfully addresses to instigate complex behaviors, include odor concentration-invariant identification, background elimination, and mixture segmentation (Hopfield 1999). From the perspective of perceptual learning in the olfaction system, the capabilities to memorize, categorize, generalize, and process odor information in a context-dependent fashion underlie its aforementioned functionality.

Despite intensive research efforts and the abundance of experimental material ranging from molecular or electrophysiological evidence to psychophysical and psychological data gathered over the past few decades, there is still a plethora of open questions and debatable hypotheses. We adopt and present a different approach to studying the mammalian olfactory system in this contribution. Namely, we attempt to understand general principles of olfactory information processing and its functional implications by constructing a computational model. The prominent role of olfactory modeling in verifying the existing and presenting novel hypotheses has been well recognized (Cleland and Linster 2005; Pearce 1997). Since we treat the system holistically, our model encompasses the first and second stages of mammalian early olfactory processing along with the first-level olfactory cortical structure, the olfactory (piriform) cortex (OC). More specifically, we synthesize olfactory stimuli patterns and simulate their processing in the olfactory epithelium (OE) in the form of olfactory receptor neuron (ORN) activations. To ensure a satisfactory level of biological plausibility of the model, the distribution of ligand–olfactory receptor (OR) affinities is generated to account for key statistical features of widely reported in vivo primary odor representations. The resulting olfactory codes are then processed in the reduced model of the olfactory bulb (OB), which constitutes an interface between the OE and the OC. In this study, OB computations are handled within the modular structure of glomerular columns, where the ORNs expressing the same OR converge and project to the corresponding group of second-order neurons. The transformation of the primary to the secondary odor representation is defined as an integration of incoming ORN activations with positive (excitatory contribution) and negative (inhibitory contribution) weights to implement a novel interval concentration coding scheme. In consequence, sparse stimulus intensity-dependent olfactory representations are produced at the output of the OB for further processing in the OC, which lies at the heart of our biomimetic approach to the odor recognition problem. The OC model is implemented in the framework of an attractor network with a hierarchical modular architecture.

The proposed three-stage abstract model of key biological mechanisms involved in olfactory information processing operates on static rate-based neural representations, which are believed to convey relevant aspects of information about odor quality and concentration (Olsen et al. 2010). Temporal features of olfactory codes and computations are therefore out of scope of this contribution. We are also interested in the scalability of the model, specifically of the OC due to its involvement in most demanding computations and the biologically plausible dimensionality of odor representations.

The study reported in this chapter is aimed at evaluating the developed abstract model of the mammalian olfactory system in test cases involving key computational tasks in odor object recognition—classification, mixture segmentation, and context-dependent identification of olfactory stimuli. The proposed mechanisms underlying the olfactory function are demonstrated in this regard. The focus here is on biomimetic aspects of these evaluation scenarios, and thus the attributes of perceptual learning and associative memory framework of the piriform cortex model, such as concentration invariance, generalization, and pattern completion, are given special attention. Our holistic modeling approach also provides insight into the impact of early olfactory coding and stimulus representations on the recognition performance. In quantitative terms, we adopt the evaluation criterion as the percentage rate of successful detection of target odor objects.

The formulated problems of odor object recognition, mainly classification and mixture segmentation, constitute a typical set of tasks addressed in machine olfaction. Despite significant differences between neural olfactory representations of natural odors and chemosensory responses to synthetic stimuli, a range of analogies in data processing and pattern analysis have been pointed out (Pearce 1997). In this study, we juxtapose the proposed biomimetic approach with classical methods in machine olfaction. One of the advantages of a biomimetic model implementing several functions in a single-system framework is the possibility to perform more than one simple task, e.g., odor discrimination, in complex scenarios more commonly encountered in natural environments. We touch upon this issue when studying a context-dependent segmentation problem. In a broader perspective, such a biomimetic network model could be combined with models of other modalities, e.g., visual, to address real-world multimodal sensory perception tasks.

The chapter is organized as follows. In the next section our biomimetic model is described in more detail. This is followed by a brief discussion of the conventional pattern recognition techniques commonly applied in machine olfaction and employed in this study for comparative purposes. The subsequent parts of the chapter are devoted to the results of evaluation of the olfactory model in test scenarios—odor object classification, mixture segmentation, and context-dependent recognition. The results are reported on the synthesized early olfactory representations and on the large-scale polymer sensor data set. The chapter concludes with a discussion of the major implications of the presented approach despite some limitations, its applicability to solve real-world odor recognition tasks in machine olfaction, and relevance to research on biological olfaction systems.