Key Words olfaction, olfactory bulb, antennal lobe, learning, dynamical systems, oscillations s Abstract We examine early olfactory processing in the vertebrate and insect olfactory systems, using a computational perspective. What transformations occur between the first and second olfactory processing stages? What are the causes and consequences of these transformations? To answer these questions, we focus on the functions of olfactory circuit structure and on the role of time in odor-evoked integrative processes. We argue that early olfactory relays are active and dynamical networks, whose actions change the format of odor-related information in very specific ways, so as to refine stimulus identification. Finally, we introduce a new theoretical framework ("winnerless competition") for the interpretation of these data.
INTRODUCTIONThe olfactory brain converts generally complex air-or water-borne chemical mixtures into singular signatures, experienced as vivid percepts. Such transformations are achieved by way of only a few brain stations-olfactory circuits are shallower than their visual and auditory counterparts-and are strongly tied to emotions and to memories acquired through other modalities. Understanding olfactory coding is thus an ambitious enterprise whose scope goes much beyond that of this review. We focus here on the sensory transformations accomplished 0147-006X/01/0301-0263$14.00 263 Annu. Rev. Neurosci. 2001.24:263-297
264LAURENT ET AL by the first two processing stages, that is, olfactory receptors and postsynaptic structures.As with any sense, understanding olfaction first requires defining the problems it has evolved to solve (Attneave 1954, Barlow 1969: segmenting an odor into its various constituents, as a chemist might do, does not appear to be one of these functions (Lawless 1997, Cain & Potts 1996. Rather, olfaction is a synthetic sense par excellence. Olfaction enables pattern learning, storage, recognition, tracking, or localization and attaches "meaning" to these patterns. By meaning we imply the richer set of associations acquired through other senses as well as hedonic (pleasant/unpleasant) and emotional valence-both of which have no physical reality outside the brain. Each one of these tasks needs to be better defined; recognition, for example, encompasses at least categorization, identification, and separation. The abilities to categorize and to identify a priori each imply very different kinds of processing; for example, categorization disregards small differences, whereas identification emphasizes them. We show how a single circuit can in fact accomplish both, through the use of dynamics.We should also exploit our understanding of the physics of odors. In vision, much attention has been given to the statistics of natural images (Field 1987(Field , 1994Olshausen & Field 1996;Rudderman 1994). Correlations across space and time make natural images highly nonstochastic. Also, the spatial-frequency ( f ) content of a natural image, be it a face or a landscape, obeys a 1/f α distributio...