Olfactory Cortex as a Model for Telencephalic Processing

views updated

Olfactory Cortex as a Model for Telencephalic Processing

Changes to myriad synapses throughout the brain must be coordinated every time a memory is established, and these synapses must be appropriately reactivated every time the memory is retrieved. Once stored, memories can be recognized (as a re-experienced input) or recalled (via different input, such as a name evoking the memory of a face or a scene evoking memories of an experience) by many routes. We remember what tables are as well as we remember a specific table, and we recognize objects despite seeing them from quite different angles, under different lighting, in different settings. Computational simulations of synaptic modifications (e.g., long-term potentiation) in distinct brain-circuit architectures illustrate how these minute changes can give rise to coherent properties of memory; how analyses of different brain areas yield derivations of disparate memory functions; and how interactions among connected regions give rise to still new operating principles beyond those of their constituents.

The principal anatomical designs in mammalian brain are cortical: planar arrays of neurons, arranged with their cell bodies in sheets and their apical dendrites standing in parallel. This laminar pattern contrasts with that of most reptilian brain structures, in which neurons are grouped in globular clusters ("nuclei"); an exception is the cortically organized reptilian pallium. The phylogenetic origins of the mammalian neocortex (perhaps including transformed nonpallial precursors as well as pallium) are the subject of continuing controversy. The difference is one of function, not just form. With cells arrayed in a plane, the axons providing input to the structure pass through the dendritic field, making random and sparse synaptic contacts. This anatomical arrangement creates a biological version of a three-dimensional array or matrix in which the rows correspond to the input axons, the columns are the dendrites, and each matrix entry is a synaptic connection between an axon and dendrite (see Figure 1). The neocortex undergoes vast expansion with mammalian evolution, and as the cortex comes to dominate the brain, cortical computation comes to dominate behavior.

The Olfactory Bulb and Paleocortex

The olfactory paleocortex, one of the oldest relics in mammals of the reptilian pallium, is an apt starting point for evaluation of cortical computation. One reason is its relative simplicity (for instance, it has three primary layers instead of the six layers of the neocortex). Another is its relative proximity to its input environment. In other sensory systems, inputs typically proceed from a peripheral organ (e.g., cochlea) to one or more lower-brain structures (e.g., cochlear nucleus, colliculus), then to a noncortical (nuclear) structure in the thalamus (e.g., medial geniculate nucleus), and only then on to the primary cortex for the appropriate sense (e.g., auditory cortex). By comparison, olfactory receptors (activated by chemical odorants drawn in through the nose) project to the olfactory bulb and thence straight to olfactory cortex. (The structure is variously termed olfactory paleocortex, for its phylogenetic age; piriform, pyriform, or prepyriform cortex, for its pearlike shape; or primary olfactory cortex, for its placement as the first cortical structure to receive olfactory input relayed from the periphery.) Abstract models have been constructed based on four fundamentals of the olfactory system: its anatomical structure, its physiological operation during behavior, the characteristics of synaptic change caused by LTP, and the nature of the inputs that arrive naturally at the system during olfactory-related behaviors.

Figure 2 schematically illustrates the anatomical structure of a typical mammal's olfactory system (adapted from Shepard, 1991). In the figure the animal's nose is to the left, with the axons from the nose comprising the first cranial nerve (Nerve I) making synaptic contact (in the regions termed glomeruli) with the primary excitatory (mitral) cells of the olfactory bulb. Mitral cells are inhibited by granule cells via specialized synapses (see Haberly and Shepard, this volume), and mitral cell axons (comprising the lateral olfactory tract) project to cortex, where they form synaptic contacts with the apical dendrites of the primary cortical excitatory layer II and III cells. Those cells in turn project both forward, to provide the input to downstream brain structures (such as entorhinal cortex), and backward, to provide feedback to the bulb both directly and by way of the anterior olfactory cortex (often termed the anterior olfactory nucleus, despite its laminar rather than nuclear structure).

Simple Emergent Computations from Feedforward Operation of the Bulb-Cortex System

When an animal is actively engaged in olfactory learning behavior, the entire bulb-cortex system, its primary target output structures (entorhinal cortex and hippocampus), and even the overt behavioral sniffing activity of the animal operate in synchrony, at a rate of about four to eight cycles per second (Macrides, 1975; Macrides et al., 1982; Vanderwolf, 1992; Wiebe and Staubli, 2001). As the animal repeatedly samples or sniffs the olfactory environment, neurons through the entire "assembly line" of olfactoryhippocampal structures send spikes down their axons, in bursts occurring approximately every fourth to eighth of a second. Computer simulations of the resulting feedforward neuronal activity in the cortex have shown that LTP-like synaptic-change increments cause specific cortical target cells that initially responded to a particular odor to become increasingly responsive not only to that odor but also to a range of similar odors. Figure 3 uses broad simplifying assumptions to illustrate this straightforward principle. (Models of the olfactory bulb [Anton et al., 1991; 1992] not discussed here are assumed).

In the left-hand panel of the diagram, input axons b, c, and d are active (arrows), and are assumed to be sufficient to elicit firing responses from three target cells (darkened). Synapses whose inputs and targets are coactive (highlighted) will potentiate. After potentiation, the enhanced synapses (enlarged, right panel) confer more voltage change than they did in their unpotentiated state, so fewer active inputs should suffice to elicit a target neuron response. Thus any of the depicted input patterns P, Q, and R may now suffice to activate the same three target cells, whereas none of these inputs would have activated these neurons prior to synaptic potentiation.

After potentiation episodes, inputs with highly overlapping activation patterns tend to educe identical neuronal response patterns in the cortex. The result is the mathematical operation of "clustering," in which sufficiently similar inputs are placed into a single category or cluster. The odor of a rose, a violet, or a lily might, after long-term potentiation, elicit only an undifferentiated response corresponding to "flower scent" (and different odors elicit only their cluster responses—e.g., meat scent, smoke scent). This cluster responsecan give rise to useful "generalization" properties, informing the organism of the category of an otherwise unfamiliar odor, but, somewhat counterintuitively, it prevents the system from making fine distinctions among members of a cluster. These results are almost generic, as many computational frameworks with very different characteristics, including competitive networks (e.g., von der Malsburg, 1973; Grossberg, 1976; Rumelhart and Zipser, 1985; Coultrip et al., 1992); backpropagation (Rumelhart et al., 1986); and dynamical or excitatory feedback networks (e.g., Hopfield, 1982) can exhibit similar properties.

Complex Computations from Combined Feedforward and Feedback Olfactory Operation

Absent from the foregoing analysis is the extensive inhibitory feedback projection from cortical neurons to granule cells in the bulb. This pathway selectively inhibits those bulb inputs that generate cluster responses in the cortex, thereby unmasking the remainder of the bulb's activity. That remainder becomes the subsequent input to the cortex on the next activity cycle, whereupon the same cortical operations are performed. The result is that the second cortical response (one fourth to one eighth of a second later) will consist of a quite distinct set of neurons from the initial response, since most of the input components giving rise to that first response are now inhibited by the feedback from cortex to bulb. Analysis of the second (and ensuing) responses has shown successive subclustering of an input: the first cycle of response identifies the odor's membership in a particular cluster (e.g., floral), the next response (a fraction of a second later) identifies its membership in a particular subcluster (rose), then in a sub-subcluster (particular variety of rose), and so on. Roughly five "levels" of subclustering occur in the simulation before the inhibitory feedback to the bulb runs its course. That is, the system uses an unexpected type of temporal coding, using specific target neurons selectively activated at a series of different time points to discriminate among inputs.

This iterative subclustering activity turned out to be mathematically expressible as a novel algorithm for the well-studied statistical task of hierarchical clustering. All such algorithms have differential costs or complexity in terms of the time (number of mathematical steps) and space (amount of storage) required for each operation. Surprisingly, the derived olfactory algorithm exhibited computational costs that compared favorably with those in the (extensive) literature on such methods (Ambros-Ingerson et al., Kilborn et al., 1996). These studies represent an instance in which a novel and efficient algorithm for a well-studied computational problem emerged from simulation and analysis of a specific cortical network. The method was readily generalized to modalities other than olfaction. For instance, input patterns corresponding to speech sounds yielded naturally occurring clusters and subclusters on successive samples (see Figure 4). Elaboration of the algorithm gave rise to families of computational signal-processing methods whose performance on complex signal classification tasks has consistently equaled or outperformed those of competing methods (interested readers are referred to: Kowtha et al., 1994; Coultrip and Granger, 1994; Granger et al., 1997).

Biological Findings and Psychological Implications

If the olfactory system operates in this way, it should show striking behavioral and electrophysiological results. Behavioral experiments showed that rats recognized novel similar odors as members of a category yet also distinguished and recognized individual category members, providing evidence that animals build unsupervised similarity-based perceptual clusters (Granger et al., 1991). Individual olfactory cortical neurons, measured chronically in behaving animals, responded selectively when tested on very different odors. Moreover, responses were transient, corresponding to the interval of a specific sniff cycle but not to multiple cycles, a result a result that also corresponds with the computer simulations (McCollum et al., 1991). Findings arrived at under different experimental conditions have yielded various hypotheses of olfactory function (e.g., Schoenbaum and Eichenbaum, 1995; Haberly, 2001). Further studies of unit neuron recordings in behaving animals will be needed to settle competing interpretations of the observed data.

The computational and neurobiological findings yield hypotheses about psychological function. Operations emerging from cortical circuits presumably constitute elemental psychological acts and contribute, through combination, to more complex mental processes in ways not yet understood. The evocation of successively finer-grained information about a stimulus via sequential cortical responses suggests a fundamental operation of repetitive perceptual sampling. Visual, auditory, and somatosensory cortices have anatomical architectures analogous to the olfactory bulb-cortex template, including excitatory feed-forward and inhibitory feedback interactions with thalamic nuclei (see Herkenham, 1986; Jones, 1998, for reviews). Perhaps the second glance of a scene educes qualitatively different information from the first glance (even when such "glances" are covert cycles operating within these cortical structures, rather than behavioral eye movements). Humans exhibit synchronized rhythmic firing during learning and during complex sensory processing (Caplan et al., 2001; Sobotka and Ringo, 1997). And human subjects in perceptual and conceptual studies robustly recognize objects first at categorical levels and subsequently at successively subordinate levels (Mervis and Rosch, 1981; Schlaghecken, 1998; Kuhl et al., 2001), suggesting the presence of structured memories that are hierarchically configured and sequentially traversed during recognition.

Modeling and analysis of other brain areas, including constituents of the hippocampal formation, auditory neocortex, the striatal complex, and thalamo-cortical loops, has yielded a range of additional, starkly different emergent fundamental computations for each structure, as well as novel complex operations from combinations of these (e.g., Lynch and Granger, 1992; Gluck and Granger, 1993; Granger et al., 1994; 1997; Myers et al., 1995; Kilborn et al., 1996; Aleksandrovsky et al., 1996; 1997). As in the case of the hierarchical clustering algorithm identified in the olfactory system, new functions derived from other brain regions exhibited computational characteristics comparable to algorithms of known power, often equaling or surpassing the best algorithms in cost and efficacy. Moreover, as in the case of the olfactory system, the results suggested new interpretations of both simple and complex psychological operations, intimating the development of more advanced hypotheses of human brain function.



Aleksandrovsky, B., Brucher, F., Lynch, G., and Granger, R. (1997). Neural network model of striatal complex. In Biological and Artificial Computation, from Neuroscience to Technology.IWANN'97 International Conference on Artificial and Natural Neural Networks. Berlin: Springer-Verlag, pp.104-115.

Aleksandrovsky, B., Whitson, J., Garzotto, A, Lynch, G., and Granger, R. (1996). An algorithm derived from thalamocortical circuitry stores and retrieves temporal sequences. Proceedings of the International Conference on Pattern Recognition, IEEE Computer Society Press 4, 550-554.

Ambros-Ingerson, J., Granger, R., and Lynch, G. (1990). Simulation of paleocortex performs hierarchical clustering. Science 247, 1,344-1,348.

Anton, P., Granger, R., and Lynch, G. (1993). Simulated dendritic spines influence reciprocal synaptic strengths and lateral inhibition in the olfactory bulb. Brain Research 628, 157-165.

Anton, P., Lynch, G., and Granger, R. (1991). Computation of frequency-to-spatial transform by olfactory bulb glomeruli. Biological Cybernetics 65, 407-414.

Bliss, T. V. P., and Lømo, T. (1973). Long-lasting potentiation of synaptic transmission in the dentate area of the anesthetized rabbit following stimulation of the perforant path. Journal of Physiology 232, 331-356.

Caplan, J., Madsen, J., Raghavachari, S., and Kahana, M. (2001). Distinct patterns of brain oscillations underlie two basic parameters of human maze learning. Journal of Neurophysiology 86, 368-380.

Coultrip, R., and Granger, R. (1994). LTP learning rules in sparse networks approximate Bayes classifiers via Parzen's method. Neural Networks 7, 463-476.

Coultrip, R., Granger, R., and Lynch, G. (1992). A cortical model of winner-take-all competition via lateral inhibition. Neural Networks 5, 47-54.

Gluck, M., and Granger, R. (1993). Computational models of the neural bases of learning and memory. Annual Review of Neuroscience 16, 667-706.

Granger, R., Staubli, U., Powers, H., Otto, T., Ambros-Ingerson, J., and Lynch, G. (1991). Behavioral tests of a prediction from a cortical network simulation. Psychological Science 2, 116-118.

Granger, R., Whitson, J., Larson, J., and Lynch, G. (1994). Non- Hebbian properties of LTP enable high-capacity encoding of temporal sequences. Proceedings of the National Academy of Sciences of the United States of America 91, 10,104-10,108.

Granger, R., Wiebe, S., Taketani, M., Ambros-Ingerson, J., and Lynch, G. (1997). Distinct memory circuits comprising the hippocampal region. Hippocampus 6, 567-578.

Grossberg, S. (1976). Adaptive pattern classification and universal recoding. I. Parallel development and coding of neural feature detectors. Biological Cybernetics 23, 121-134.

Haberly, L. (2001). Parallel-distributed processing in olfactory cortex, New insights from morphological and physiological analysis of neuronal circuitry. Chemical Senses 26, 551-576.

Herkenham, M. (1986). New perspectives on the organization and evolution of nonspecific thalamocortical projections. In E. Jones and A. Peters, eds., Cerebral cortex, Vol. 5. New York: Plenum Press.

Hopfield, J. (1982). Neural networks and physical systems with emergent collective computational abilities. Proceedings of the National Academy of Sciences of the United States of America 79, 2,554-2,558.

Jones, E. G.(1998). A new view of specific and nonspecific thalamocortical connections. Advances in Neurology 77, 49-71.

Karten, H. J. (1997). Evolutionary developmental biology meets the brain, the origins of mammalian neocortex. Proceedings of the National Academy of Sciences of the United States of America 94, 2,800-2,804.

Kilborn, K., Granger, R., and Lynch, G. (1996). Effects of LTP on response selectivity of simulated cortical neurons. Journal of Cognitive Neuroscience 8, 338-353.

Kowtha, V., Satyanarayana, P., Granger, R., and Stenger, D. (1994). Learning and classification in a noisy environment by a simulated cortical network. Proceedings of the Third Annual Computer and Neural Systems Conference Boston: Kluwer.

Kuhl, P., Tsao, F., Zhang, Y., DeBoer, B. (2001). Language, culture, mind, brain, Progress at the margins between disciplines. Annals of the New York Academy of Science 935, 136-174.

Lynch, G., and Granger, R. (1992). Variations in synaptic plasticity and types of memory in cortico-hippocampal networks. Journal of Cognitive Neuroscience 4, 189-199.

Macrides, F. (1975). Temporal relations between hippocampal slow waves and exploratory sniffing in hamsters. Behavioral Biology 14, 295-308.

Macrides, F., Eichenbaum, H. B., and Forbes, W. B. (1982). Temporal relationship between sniffing and the limbic (theta) rhythm during odor discrimination reversal learning. Journal of Neuroscience 2, 1,705-1,717.

McCollum, J., Larson, J., Otto, T., Schottler, F. Granger, R., and Lynch, G. (1991). Short-latency single-unit processing in olfactory cortex. Journal of Cognitive Neuroscience 3, 293-299.

Mervis, C., and Rosch, E. (1981). Categorization of natural objects. Annual Review of Psychology 32, 89-115.

Myers, C., Gluck, M., and Granger, R. (1995). Dissociation of hippocampal and entorhinal function in associative learning, A computational approach. Psychobiology 23, 116-138.

Puelles, L. (2001). Thoughts on the development, structure and evolution of the mammalian and avian telencephalic pallium. Philosophical Transactions of the Royal Society of London 356, (Biol.) 1,583-1,598.

Rosch, E., and Lloyd, B. B. (1978). Cognition and categorization. Hillsdale, NJ: Erlbaum.

Rumelhart, D., Hinton, G., and Williams, R. (1986). Learning representations by back-propagating errors. Nature 323, 533-536.

Rumelhart, D., and Zipser, D. (1985). Feature discovery by competitive learning. Cognitive Science 9, 75-112

Schlaghecken, F. (1998). On processing "beasts" and "birds," an event-related potential study on the representation of taxonomic structure. Brain and Language 64, 53-82.

Schoenbaum, G., and Eichenbaum, H. (1995). Information coding in the rodent prefrontal cortex. I. Single-neuron activity in orbitofrontal cortex compared with that in pyriform cortex. Journal of Neurophysiology 74, 733-750.

Shepherd, G. (1991). Computational structure of the olfactory system. In J. L. Davis and H. Eichenbaum, eds., Olfaction, a model system for computational neuroscience. Cambridge: MIT Press.

Sobotka, S., and Ringo, J. (1997). Saccadic eye movements, even in darkness, generate event-related potentials recorded in medial septum and medial temporal cortex. Brain Research 756, 168-173.

Vanderwolf, C. (1992). Hippocampal activity, olfaction, and sniffing: An olfactory input to the dentate gyrus. Brain Research 593, 197-208.

von der Malsburg, C. (1973). Self-organization of orientation sensitive cells in the striate cortex. Kybernetik 14, 85-100.

Wiebe, S., and Staubli, U. (2001). Recognition memory correlates of hippocampal theta cells. Journal of Neuroscience 21, 3,955-3,957.

Richard H.Granger, Jr.