Workshop 4: Olfaction: Abstracts and Lecture Materials
Locating the source of odor in a turbulent environment --- a common behavior for living organisms --- is non-trivial because of the random nature of mixing. We analyze the statistical physics aspects of the problem and propose an efficient strategy for olfactory search which can work in turbulent plumes. The algorithm combines the maximum likelihood inference of the source position with an active search. Our approach provides the theoretical basis for the design of olfactory robots and the quantitative tools for the analysis of the observed olfactory search behavior of living creatures (e.g. odor modulated optomotor anemotaxis of moth).
Intracellular recordings in vivo from locust antennal lobe (AL) projection neurons (PNs) revealed that individual PNs phase-lock with population oscillations at times that depend on the stimulus. GABAergic input from local AL neurons provides a potential mechanism for PN synchronization and the transient nature of PN synchronization is linked to variations in inhibitory drive over the duration of a response (Bazhenov et al., Neuron 30:553-567, 2001; 30:569-581, 2001). Kenyon cells (KCs) of the mushroom body (MB) - postsynaptic target of PN afferents - decode stimulus-specific spatio-temporal patterns of PN activity. In vivo recordings from the locust MB demonstrate high specificity in Kenyon cells' (KCs) responses during odor processing (Perez-Orive et al., Science 297:359-365, 2002). With computer models of the locust olfactory system we explored intrinsic and synaptic mechanisms involved into decoding spatio-temporal patterns of AL activity in the MB. In the model, a combination of the KCs' active conductances and inhibitory feedback from lateral horn interneurons provided a potential mechanism to detect synchrony in the input spike trains from the AL. Inputs with noise were decoded more reliably when KC firing was controlled by spikes' synchrony (coincidence detection) rather than by rate of spikes (integration). Even with the intrinsic (active conductances) and synaptic (input from lateral horn) mechanisms present, synchrony detection by the KCs would be impossible without initial tuning of the synaptic weights between the AL and the MB. We proposed that spike-timing dependent plasticity (STDP) operating on the synaptic afferents between AL and MB can provide the necessary mechanism. In the model, the trained network responded with the same pattern of spiking KCs, which usually included only 1-3 "odor-specific" neurons, independently of the initial (before training) synaptic weights. Our results suggest that STDP-based learning rule operating on synaptic afferents between AL and MB can control and tune synaptic weights to provide high specificity of KCs' responses to presented odors, a process that might play a role at early stages of development. This tuning combined with nonlinear intrinsic properties of the KCs and inhibitory circuits of the MB can allow MB interneurons to operate as coincidence detectors selecting correlations in the input spike trains.
The spatial patterning of odor-evoked glomerular activity in the vertebrate olfactory bulb and arthropod antennal lobe reflects an important component of the first-order olfactory representation and contributes to odorant identification. Higher concentration odor stimuli evoke broader glomerular activation patterns, resulting in greater spatial overlap among different odor representations. However, behavioral studies demonstrate results contrary to what these data suggest: both mice and honeybees are more, not less, able to discriminate among odorants presented at higher concentrations. Using a computational model incorporating the honeybee antennal lobe and an identified reward interneuron, and utilizing spike timing-dependent synaptic plasticity (STDP), a hebbian learning rule derived from physiological studies in mammalian hippocampus, we show that changes in spike synchronization patterns among projection neurons, as observed electrophysiologically, could parsimoniously underlie these observations. Existing physiological and psychophysical data suggest that this model may be appropriately extended to the vertebrate olfactory bulb as well. The results suggest that effective activity among projection neurons, defined as spiking activity that influences the firing of a postsynaptic neuron, may be considerably sparser than overall spiking activity, and that stimulus salience, as defined behaviorally, is modified by stimulus intensity and directly correlated with the degree of spike synchronization among second-order olfactory neurons.
Neural circuits that process inputs from sensory organs, such as those in the olfactory bulb (OB), must detect the presence of a stimulus and differentiate one stimulus from another. To achieve this they need to have mechanisms to selectively emphasize salient features of the input stimulus, in other words, to increase signal to noise and to enhance ?contrast? between different stimuli. Such computations by neural circuits depend upon the intrinsic electro-physiological characteristics of neurons, the properties of the synapses that link them and the interaction between these cellular and synaptic properties In granule cells we have found a calcium activated non-specific current that can be induced by action potential activity or large amplitude excitatory synaptic currents that are capable of elevating cytoplasmic [Ca2+]. It can thereby amplify large but not small synaptic potentials. Furthermore, the channel?s sensitivity to Ca2+ is potentiated for many seconds after a brief period of [Ca2+] elevation. Thus a burst of action potentials that precede a synaptic input by several seconds will enhance the synaptic potential provided the potential is large enough to generate a Ca2+ influx itself. In mitral cells we have found dopaminergic D2-type receptors are capable of inhibiting transmitter release from the basal dendrites. This results in reduced inhibitory feedback from granule cells/ A portion of the inhibition of release arises from a activation of a transient K+ current that causes attenuation of the propagation of action potentials into the dendrites resulting from D2-receptor mediated hyperpolarization. Consequences of this offsetting combination of reduced excitability due to hyperpolarization and disinhibition resulting from reduced excitation of granule cells will be presented and discussed.
In this talk I will describe several different approaches to the modeling of oscillations and waves found in the procerebral lobe of the terrestrial mollusc, Limax. I will start with a very simple model, laying out several possible mechanisms for waves and show how experimental results select one of them. Next I describe some consequences of waves and synchrony for the learning of new odors. Here I cite some rather surprising experiments which show that certain groups of neurons are selectively stained only after learning. Finally I close with some detailed biophysical models which explain the role of gap junctions and nitric oxide in the synchronization of rhythms in the PC lobe.
The olfactory system performs remarkable feats of molecular detection and pattern recognition using what are, in engineering terms, very imprecise computing elements (neurons). Olfactory systems are comprised of an array of odor sensors providing input to a computational engine that performs pattern recognition and sensory integration on the input pattern. By using large numbers of each type of sensory neuron and steadily replacing and renewing each individual sensor, biological olfactory systems have evolved odor sensing and analysis capabilities unmatched by current technology (Gelperin & Hopfield, 2002). The olfactory system of the terrestrial slug Limax maximus provides a convenient experimental system in which to address general issues of olfactory computation. The olfactory system of Limax incorporates many of the general design features of the mammalian olfactory system (Hildebrand & Shepherd, 1997), including coherent oscillatory activity dependent on nitric oxide (Gelperin, 1999). Synaptic plasticity and learning are essential components of olfactory processing (Hudson, 1999) and involve the earliest stages of central synaptic processing of olfactory input (Ermentrout et al., 2001; Yuan, Harley & McLean, 2003).
An early model of learning in the Limax olfactory system (Gelperin, Hopfield & Tank, 1986) delt with central odor representations that could mediate several forms of higher-order conditioning exhibited in behavioral experiments with Limax. More recent work has incorporated the wave propagation and band-like odor representations found in this system [Ermentrout, 2001 #927}. Recent extensions involving adult neurogenesis of olfactory interneurons and learning-activated gene expression in the Limax olfactory centers will be described.
- Ermentrout, B., Wang, J. W., Flores, J., & Gelperin, A. (2001). Model for olfactory discrimination and learning in Limax procerebrum incorporating oscillatory dynamics and wave propagation. J Neurophysiol, 85, 1444-1452.
- Gelperin, A. (1999). Oscillatory dynamics and information processing in olfactory systems. J Exp Biol, 202, 1855-1864.
- Gelperin, A., & Hopfield, J. J. (2002). Electronic and computational olfaction. In P. Given, ed., Chemistry of taste. Washington, DC: American Chemical Society, pp. 289-317.
- Gelperin, A., Hopfield, J. J., & Tank, D. W. (1986). The logic of Limax learning. In A. I. Selverston, ed., Model neural networks and behavior. New York: Plenum Press, pp. 237 - 261.
- Hilderbrand, J. G., & Shepherd, G. M. (1997). Mechanisms of olfactory discrimination: Converging evidence for common principles across phyla. Ann Rev Neurosci, 20, 595-631.
- Hudson, R. (1999). From molecule to mind: the role of experience in shaping olfactory function. J Comp Physiol A, 185, 297-304.
- Yuan, Q., Harley, C. W. & McLean, J. H. (2003). Mitral cell beta(1) and 5-HT(2A) receptor colocalization and cAMP coregulation: A new model of norepinephrine-induced learning in the olfactory bulb. Learn Mem, 10, 5-15.
Olfactory computation by synchrony in a spike timing synchronization across different neurons can be selective for the situation where the neurons are driven at similar firing rates, a "many are equals" computation. Based on this principle, we instantiate an abstract algorithm for robust odor recognition into a model network of spiking neurons whose main features are taken from properties of the primary stages of biological olfactory systems, and demonstrate its computational abilities in tasks such as intensity-independent recognition and background rejection. Synaptic plasticity enables a neural system to 'learn' or 'adapt'. We derive a plasticity rule based on the idea that synapse change must also be capable of keeping a network in a functional state by repairing the damage that spontaneously occurs. The derived timing-based plasticity rule also makes the system capable of learning new odors in an unsupervised fashion. The rule is strikingly similar to those experimentally found in LTP.
The mammalian olfactory bulb (OB) is a unique cortical structure as it receives input directly from peripheral olfactory receptor neurons. It is also unique as a first sensory relay in that it receives massive centrifugal input from central brain areas, primarily from other parts of the olfactory system and nearly every part of the limbic system. Single mitral cells receive receptor neuron input and project out of the OB. Inhibitory interneurons (granule cells) receive much of the centrifugal input and form lateral connections with mitral cells, and putatively with each other. I will discuss behavioral-state dependent activity and activity changes at the single cell and population level. Activity of mitral cells, as measured by firing rates during operant behavior, is modulated strongly by the behavioral association of an odorant. OB gamma oscillations (40-100 Hz), which have been shown to be associated with odor sniffing in some cases are also dependent on behavioral state and change character dramatically dependent on the type of behavior (waiting, alert immobility, odor discrimination, exploratory sniffing, etc.). The gamma band is in fact two behaviorally distinct bands, and preliminary studies from knockout mice show that the two types of oscillations, termed type 1 and type 2 gamma, may be dependent on different types of synaptic interactions and behavioral states.
Olfactory systems in biological organisms are capable of performing amazing feats of molecular identification and analysis. These systems manage to overcome the limitations of imprecise neuronal hardware and identify components in complex mixtures over twelve orders of magnitude in concentration. Is it possible to replicate some of these sensory and computational capabilities in artificial systems? I will describe some computational algorithms needed for artificial olfactory systems to approach some of the capabilities of biological organisms. I will also demonstrate a mobile robotic olfactory system that is equipped with electronic chemical sensors, otherwise known as an ``E-nose.'' The challenge for computational algorithms is to obtain the maximal amount of information about the external olfactory environment from the sensor readings, without being corrupted by noise. I will describe recently developed learning algorithms for dimensionality reduction that are able to process the data from the sensors, and to recognize the different patterns of activity.
Based on experiments with olfactory network dynamics, Clione hunting activity and generation of neural sequence in a songbird, we discuss the dynamical principles that underlie the sequential activity of complex neural ensembles. It looks like context or stimuli dependent sequential activity of the groups of neurons that are able to organize the sequential order in thoughts and actions is one of the most powerful ideas that nature uses in nervous systems. We think that the Winnerless Competition (WLC) principle is controlled such neural activity.The main point of this principle is the transformation of the incoming identity code or spatial code into ensemble (spatio)-temporal output based on the intrinsic switching dynamics of the neural network. In the presence of stimuli, the sequence of the switching, whose geometrical image in phase space is a heteroclinic contour, uniquely depends on the incoming information. Together with the results of computer modeling of the networks with different levels of complexity, we present rigorous results about the existence and stability of the WLC dynamics and discuss the advantages of sequential coding of neural information.
Animals modify responses to both pheromones and to nonpheromonal odors based on experience. The behavioral mechanisms through which responses become modified depend on a number of factors. The neural systems that underlie olfactory processing in phylogenetically diverse groups of organisms appear to have evolved similar means to accomplish olfactory coding. Dozens to thousands of somewhat broadly tuned receptors set up a combinatorial code for odors. Yet recent evidence indicates that neural plasticity occurs at early stages of sensory processing. I propose that this plasticity is used to help early processing extract important features from a variable olfactory environment in which targeted odor "objects" differ from time-to-time and place-to-place. My laboratory has identified several behavioral mechanisms (synthetic/configural processing, blocking and attention-like components) used to analyze odor objects. And I propose that neural mechanisms that underlie this behavioral plasticity are those that are set up in early sensory processing. These hypotheses are evaluated using several different insect species, each of which can be used to approach these issues from different combinations of behavioral, neurophysiological and molecular perspectives. The presentation will summarize information about olfactory plasticity from these species. But the point will also be made that we still lack a detailed understanding of the "scene statistics" of odor objects. Any thorough understanding of behavioral and neural mechanisms of olfactory plasticity cannot be complete without a knowledge of how natural odor objects vary from time-to-time and place-to-place.
We examined the encoding and decoding of odor identity and concentration by neurons in the first and second relays of the locust olfactory system, the functional analogues of the vertebrate olfactory bulb and olfactory cortex. We found that both the identity and concentration of odorants are represented as spatiotemporal patterns of activity across projection neurons (PNs, analogous to mitral cells) in the first relay. The distributed activity patterns can be classified by similarity as concentration groups contained within larger odor identity clusters. When analyzed as functions of time, the high-dimensional PN activity patterns can be described as families of trajectories lying on low-dimensional manifolds. Each manifold represents odor identity and contains trajectories representing different concentrations. The spatiotemporal activity patterns are then decoded piecewise over very short time periods and compressed into highly informative responses about both identity and concentration by downstream neurons in the second relay.Some references:
- Laurent, G. (2002, November 3). Olfactory network dynamics and the coding of multidimensional signals. Nat Rev Neurosci., 11, 884-95.
- Perez-Orive, J., Mazor, O., Turner, G., Cassenaer, S., Wilson, R., & Laurent, G. (2002). Oscillations and sparsening of odor representations in the mushroom body. Science, 297, 359-365.
- Laurent, G., Stopfer, M., Friedrich, R., Rabinovich, M.I., Volkovskii, A., & Abarbanel, H.D.I. (2001). Odor encoding as an active, dynamical process: Experiments, computation and theory. Annu. Rev. Neurosci., 24, 263-297.
