The Physiological Basis of Mental Images
Biological Psychiatry, Vol 18, No. 10, 1983

Academic Address


The Physiological Basis of Mental Images

Walter J. Freeman

Presented at the Annual Meeting of the Society of Biological Psychiatry, New York April 29, 1983

Supported by grants from NIMH (MH06686) and NIH (NS16559). Department of Physiology-Anatomy, University of California, Berkeley, California.


An important hypothesis in neurophysiology supposes that sensory information is encoded in spatial patterns of neural activity. A test of this hypothesis in rabbits reveals that the spatial pattern of encephalographic (EEG) activity of the olfactory bulb depends less on odor than on expectation of odor. Analysis of the structure and nonlinear dynamics of the olfactory system suggests that the EEG manifests a neural activity pattern that is an active filter corresponding to what ethologists call a "chemical search image. " This first glimpse at the physical aspect of what may be a primitive form of mental image allows us to infer some of its properties. It is self-organized; input destabilizes the bulb, and the event runs a course that is shaped by the initial conditions, the sensory input, the chemical state of the bulb from centrifugal input, and the record of past input embedded in synaptic connections. The event is analog, but not an analog of input. Its intensity pattern exists in two dimensions but it is not a picture. Its bases in feedback and resonance make it connotative and ascriptive, not enotative or descriptive. The neural-mental image is an operator, not an operand It gathers neurons by the millions into coherent activity and creates information, which it transmits to the next stage in the brain, thereby helping to shape behavior. It cannot be understood without reference to both what the subject does and what is done to it. Its function in the bulb is to regulate olfactory input with minimal and nonspecific centrifugal control.



     The phrase "information processing" has come into common use over the past 20 years to signify transformation of sensory input to motor output by neural operations. It is a metaphor from computer technology. Inasmuch as brains do not work like computers either physically or conceptually, there is some need to examine its meaning with respect to what physiologists actually do when we claim to study it.


     In the simplest sense information is conveyed by a sensory stimulus that gives rise to a behavioral response; by exclusion the information can be identified with some attribute of the stimulus. It must alter the activity of sensory receptors, which causes a succession of changes in brain activity. The prime task of the physiologist is to observe the sensory input and the patterns of neural activity at a sequence of stages through the brain. At each stage the comparison of the output with the input defines an operation. Each input and output pattern must be described with numbers or an equivalent geometry, so that the comparison has the form of a mathematical operation or equation. The physiology of information processing is described by an ordered arrangement of the mathematical operations between the measured stimulus and the measured response.

     More particularly, we look for the transformations of meaning and not merely the transformation of activity patterns. Not all sensory stimuli convey information, and not all neural operations reflect processing. We can reasonably construe that the vastly greater part of sensory input never achieves informational status by virtue of neural mechanisms of attention and habituation that exclude undesired stimuli. Much of brain activity reflects merely its basal or null state and the operations of homeostasis that maintain it. Even in the study of a conditioned reflex that provides a well-defined context of goal-directed behavior in which information is surely being processed, we have no guarantee that what we can measure will instantiate information in the stimulus or the neural activity patterns. Nevertheless, if we understand the anatomy and physiology of the nervous system in sufficient depth we can construct some useful hypotheses. My approach in this lecture is to sketch the structure and dynamics of one sensory system, describe some measurements of its neural activity relating to learned behavior, construct a hypothesis on a primitive process of generalization over equivalent stimuli, and discuss some of its implications about the nature of brain function. In particular I hope to outline a neural mechanism for mental images.



     I have chosen to study olfaction because it is the simplest sensory system in the context of goal-directed behavior. It consists of three structures in series: the receptor array, the bulb, and the cortex (Fig. 1). Each receptor has a ciliated border extending into the nasal cavity for odor reception and an unbranched axon that extends in the primary nerve to the bulb. There it synapses onto the dendrites of a mitral cell, which is the principal projection neuron of the bulb. They in turn send their axons in the olfactory tract to the olfactory cortex to form synapses on the dendrites of superficial pyramidal cells. The deep pyramidal cells distribute their axons broadly to the thalamic, limbic, and motor systems of the brain, including the amygdalostriatum, hippocampus, pallidum, hypothalamus, and reticular formation. It is apparent that chemotransduction takes place in the receptor layer, and that the information processing which is peculiar to odor identification takes place within the olfactory bulb and cortex.


Fig. 1. The block diagram shows the three main stages of the olfactory system. The main cell types are schematized in the center. The main flow of activity is inward (centripetal) to the brain, but there is feedback within the bulb and cortex and from the brain to both by centrifugal pathways.


     Three topological features of this system require further comment. First, the main flow of olfactory information is forward, but the main work of the system is done by prominent feedback pathways both within and between the bulb and cortex. These are provided by interneurons and by axon collaterals between projection neurons. There is also feedback from the limbic and motor systems, including the brain stem, which provides for centrifugal controls of the bulb and cortex. The cranial nerve and respiratory nuclei provide for a larger feedback system that controls air flow and the delivery of odors to the receptors.

     Second, the numbers of neurons and connections in these structures are exceedingly high. For example, in the cat there are on the order of 10^8 receptors, 105 bulbar projection neurons, 10^8 bulbar interneurons, and 10^7 to 10^8 cortical neurons. There are no known feedback pathways among receptors, but there are very high densities of synaptic interconnection among bulbar and cortical neurons (Fig. 2). The best way to view the function of the system is in terms of three layers of neurons, in which events exist as spatial patterns of activity in the two dimensions of the surface of each layer. Transmission is simultaneous and in parallel from one surface to the next, and transformations take place laterally within each layer.

     Third, there is a measurable degree of topographic order in the projection of axons from the receptors to the bulb, but far less from bulb to cortex. Axons from each local region of the bulb diverge broadly but unevenly over the cortex, and conversely each local region of the cortex receives from a broad distribution of bulbar neurons. In effect, the primary pathway preserves a degree of point-to-point representation from receptors to bulbar neurons, whereas the tract provides the anatomical basis for a spatial integration by cortical neurons over bulbar output. At this stage topographically specific information is deleted, as it must be in all sensory systems after information preprocessing. Roughly speaking we can compare the system to a camera, in which the array and nerve act like a filter and the tract acts more like a lens.



Fig. 2. Each stage is formed by a sheet of neural tissue. Activity is described in the two surface dimensions. Transformation is based on multiple types of feedback within the bulb and cortex.



     The electrical activity of the system occurs in two forms. One is in the form of pulses or units generated by axons; the other is in the form of electrotonic currents that are manifested in field potentials generated by dendrites and cell bodies. They are powerfully determined by respiration. With each inspiration there is a wave of receptor potential called the electro-olfactogram that manifests the excitation of receptors. This is accompanied by a surge of pulses in the primary nerve. The bulb generates a similar slow wave of electrical potential that manifests its excitation. Superimposed on the wave is a burst of oscillatory potential at 40 to 80 Hz that manifests alternating excitation and inhibition of bulbar neurons owing to the negative feedback connections. These waves constitute the electroencephalogram (EEG) of the bulb (Fig. 3). Bulbar neurons also show a surge in pulse activity with inspiration. Statistical analysis shows that the probability of firing of each neuron oscillates at the same frequency as the EEG potential. These probabilities sum by virtue of the spatial integration in bulbocortical transmission over the tract, so that each local neighborhood in the bulb can be said to generate a continuous pulse density (pulses/second/unit area of an ensemble in the bulb), owing to the high density of neurons (Fig. 4). This pulse density wave drives the cortex, so that the oscillation occurs in the cortical EEG and pulse density waves at the same frequencies as in the bulb.


Fig. 3. Neurons and nerve cell assemblies generate activity in two modes: the wave and the pulse. With each inspiration there is a smooth wave of receptor activity and a surge of pulses in the nerve. The bulb generates an oscillatory wave called a burst of EEC activity. This is accompanied by an increase in the density of pulse activity in the bulb.


N O R M A L I Z E D   C O N D I T I O N A L

P R O B A B I L I TY   S U R F A C E

Fig. 4. An example is shown of the normalized statistical relation for a single cell of mitral cell pulse probability (vertical axis) conditional on the EEG amplitude (right upward axis) at various time lags from the EEG sample (horizontal axis). The time-dependent oscillation is at the frequency of the EEG (Freeman, 1975).


     Generally, forward transmission is by excitation. In the main the receptors are selectively excited by odors, their axons excite mitral cells, and they in turn excite cortical neurons. The mitral and pyramidal cells are excitatory to each other and to the interneurons. The principal interneurons in the bulb and cortex are inhibitory to the projection neurons and to each other. Thereby three types of feedback arise: negative feedback between excitatory and inhibitory cells, and two forms of positive feedback, namely, mutual excitation and mutual inhibition. The feedback is the basis for interaction, and this creates ensembles that are the functional neural units for information processing.



     Thus far I have discussed the brain in the way that a computer technician might describe the circuit diagrams and functional properties of some components in an information processing device. For the study of information processing the device must be fed some information to see what it does with it. This is done in the olfactory system by giving the receptors an odor and observing the neural activity. In order to describe the relevant attributes of neural activity we require a hypothesis or search image of the way in which the olfactory system encodes its information.

     The most widely held view is that olfactory coding is spatial. This hypothesis is based on the results of numerous investigators in recording the pulse activity of single receptors and of neurons in the bulb and cortex, which show that individual receptors and neurons respond differentially to selected odors, although with extensive overlap in response profiles of neurons to a battery of test odors. Accordingly, for each discriminable odor there is a subset of receptors that can respond to that odor, and these receptors have a pattern of spatial location in the nasal cavity that is unlike the pattern for any other subset for another odor. Odor intensity is thought to be conveyed in the number of receptors activated and in their rates of firing. By the law of specific nerve energies the unique signature established in the primary nerve by a given odor must derive from the spatial pattern of receptor activity evoked by that odor. Because there is a topographic order in the projection of receptor axons to the bulb, there should also exist a characteristic spatial pattern of neural activity for that odor over the surface of the olfactory bulb, which differs from the receptor activity pattern but is still unique to that odor.

     An ideal physiological test of the spatial coding hypothesis would be to record simultaneously during one inspiration from receptors and bulbar neurons in sufficiently large numbers to describe the spatial patterns of activity in both the nose, the bulb, and the tract. Comparison of the three patterns would manifest directly the neural operations performed by the primary nerve and the bulb on receptor output to give bulbar output. This approach is not feasible, because it is technically not possible to record pulses simultaneously from a sufficiently large sample of neurons to define their spatial patterns of activity.

     A minimally sufficient test is to record the wave activity at the surface of the bulb during stimulation of the receptors with test odors. According to the hypothesis each odor should give rise to a spatial pattern of EEG activity that is unique to that odor. This test was performed on cats and rabbits with arrays of 64 electrodes surgically implanted onto the surface of the bulb or cortex. After recovery from surgery the subjects were presented with various odors, and the spatial patterns of the EEG were observed before and during odor presentation. No dependence was found of EEG spatial pattern on the odors presented (Freeman, 1978). The amplitude of the EEG burst was uneven over the bulb. Contour plots of the amplitude resembled maps of mountains rising from a plain. Each subject revealed its characteristic pattern that, like a signature, was easily recognizable but never twice the same. The variation between burst patterns with and without odors was no greater than the variation among sequential bursts without odor.

     However, in that test there was no behavioral confirmation that olfactory information was being processed. The proper condition was established through training rabbits to respond to an odor as a conditioned stimulus (CS) by pairing the odor with a brief shock as an unconditioned stimulus (US). On the first day of training a conditioned response (CR) developed rapidly (Davis and Freeman, 1982), and there was a difference between the EEG pattern during odor presentation compared with the prestimulus pattern. In subsequent sessions the CR persisted but a new spatial pattern stabilized and was present whether or not the odor was present. With a new odor CS the CR transferred to the new CS, and a new EEG pattern difference appeared during the odor period as compared with the preodor period. Thereafter a new pattern stabilized. Such EEG changes did not occur upon presentation of odors without reinforcement nor on presentation of visual or auditory stimuli with reinforcement. Over numerous sessions and subjects we concluded that changes in the EEG took place with changes in expectancy of odors and not merely with their presentation (Freeman and Schneider, 1982). By a "change in expectancy" we meant a change in disposition to act in response to an odor. That is, neural activity at this first synapse in the bulb had to be described with respect to response as well as to stimulus.

     These results are in striking accord with observations by Lancet et al. (1982) and others who administered radiolabeled 2-deoxyglucose to rabbits during the sustained (45 min) presentation of an odor. The bulbar regions of enhanced metabolic activity conform in size, shape, and location to the active foci revealed by EEG recording. However, the labeling method allows use of only one odor or combination of odors for each animal. Considering the variability between subjects in the basal EEG pattern, that method does not allow one to distinguish the dependence of spatial pattern of activity on odor versus expectancy.

     One of three hypotheses must follow this outcome. First, the spatial coding hypothesis may be invalid. This is most unlikely because of a large body of converging evidence that I need not review here. Second, the EEG may not manifest spatially coded olfactory information. The changes in EEG with conditioning might result epiphenomenally from a variety of possible nonspecific or trophic changes in the bulb. We have not entirely excluded these possibilities, but our present evidence makes them seem unlikely. Third, the information that the EEG does manifest may not be from the input but from something else. My present view is that the EEG manifests the activity of a nerve cell assembly as first defined by Donald Hebb (1949). The critical question is: What is the neural mechanism of burst formation?



     The EEG and pulse activity are closely covariant temporally and spatially because dendritic current density (related to EEG wave amplitude) modulates pulse rates (related to pulse densities), and pulse rates modulate dendritic currents. Each neuronal assembly performs two operations, one at synapses where incoming axons end on dendrites, and the other at the trigger zone where the dendrites converge and the axon begins. Conversion of pulses to waves at synapses is constrained to a small linear range, so that wave amplitude is proportional to pulse density input. By definition the proportionality coefficient represents the synaptic gain. It is called the gain because it is a number by which the axonal input is multiplied to give the dendritic output. It is subject to increase or decrease with learning or habituation (Freeman, 1975; 1979b).

     The conversion of wave amplitude to pulse density at trigger zones is nonlinear; experimental data show that the relationship is S-shaped or sigmoidal. With excessive inhibitory or excitatory wave action it approaches, respectively, zero or a plateau maximum. I measured the relation by fitting a curve to the data with an equation I derived in part from the Hodgkin-Huxley equations (Fig. 5). The slope of the sigmoid curve represents the strength of axonal output for a given dendritic input to the axons; it is called the axonal gain. In contrast to the synaptic gain, the axonal gain increases with increasing arousal; the entire curve increases in height without change in basic form, so that the aspect of axonal gain relating to arousal can be represented also by a number. The overall forward gain for each set of neurons is given by the product of the synaptic gain and the axonal gain, because the two operations are in series. It can be increased either at the synapse by factors relating to learning or at the axon by factors relating to arousal.

     The maximal axonal gain is displaced to the excitatory side of the rest gain for zero wave amplitude. This is a crucial finding. Herein lies the explanation for the occurrence of the bulbar oscillatory burst with inspiration. Receptor input excites mitral cells, which excite each other through axon collaterals. Owing to the nonlinear wave-pulse conversion there is a coupled increase in excitatory activity and in excitatory forward gain and hence feedback gain. This also increases the negative feedback gain, so that when some activity threshold is exceeded, the projection and interneurons break into a sustained oscillation that lasts for the duration of the surge in receptor input (Freeman, 1979b).

     This property is of fundamental importance for understanding the function of the bulb. Input to the bulb does not simply impose activity upon it in a form determined by the input. The input serves to destabilize the bulb locally so that it goes beyond receptor control. The response is determined in part by the initial conditions imposed by the input but in larger part by the state of the intrinsic synaptic connections of the bulb. This quasi-independence is manifested most obviously in the rhythmic character of the bulbar response to the slow surge of receptor input.


Fig. 5. Three examples are shown above of the curve that was tatted to the statistical data relating wave amplitude v to pulse density p. Below are shown three examples of the slopes of the three curves, dp/dv, showing the type of change with arousal. The height of the slope is a measure of the gain of this operation, The maximal gain is to the excitatory side of resting or prestimulus wave amplitude at the triangles. Inhibition is to the left, excitation to the right (Freeman, 1979a).

     These dynamic properties have been represented and analyzed by means of coupled nonlinear integrodifferential equations called a KII model. The time and space constants were evaluated in animals under deep anesthesia (Freeman, 1975). The normative values of the coefficients representing interactive gains were evaluated by measurement of averaged evoked potentials and poststimulus time histograms derived with the use of electrical stimulation. The solutions of the equations that simulate EEG activity have shown that the gain for mutual excitation plays a particularly important role in determining the spatial patterning of bulbar activity. It is an extremely sensitive parameter. Increases in simulated synaptic strength on the order of 40% can increase local sensitivity on the order of 40,000-fold owing to the combination of the nonlinear function and the mutually excitatory connections (Freeman, 1979c). According to the KII model each local neighborhood of the bulb has its own value of gain. If the gain values among local subsets are spatially uniform, then the spatial pattern of output activity is determined by the spatial pattern of input. If the excitatory gains in some neighborhoods are increased, then the spatial pattern of output activity tends to conform to the spatial pattern of the more strongly interconnected subsets and not to input. I propose that increases in synaptic gain occur during learning, thereby forming a template.



     My line of reasoning is as follows. The EEG changes are specifically in burst amplitude, so whatever mechanism is postulated must explain that attribute. The high coherence of the EEG burst, particularly its common frequency, indicates that bulbar neurons are widely interactive, particularly by way of mutually excitatory synaptic connections among bulbar excitatory neurons, the mitral cells. This type of synapse has already been shown to increase in strength during associative learning. In an earlier evoked potential study cats were trained to respond to an electrical stimulus to the olfactory tract as a CS. When the subjects acquired a CR to the electrical stimulus, the shape of the averaged evoked potential for the olfactory cortex changed in a manner unique to the induction of selective attention to the CS (Emery and Freeman, 1969). That unique configuration of change (Fig. 6) was simulated by the same KII model used for the EEG, and only by an increase in mutually excitatory gain (Emery and Freeman, 1969; Freeman, 1979b). The same configuration of change was seen in the antidromically evoked potential in the olfactory bulb (unpublished data).

     Further evidence was obtained (Freeman and Schneider, 1982) from rabbits trained to respond to an odor and not to an electrical stimulus given to the tract. In naive subjects the spatial pattern of amplitude of the oscillatory component of the bulbar antidromically evoked potential conformed to the spatial pattern of the EEG induced by orthodromic receptor input. When the subjects were trained to respond to an odor, the EEG pattern changed, and so also in conformance did the pattern of the evoked potential, showing that the spatial pattern of both was determined by patterns of synaptic connection strength not between receptor axons and bulbar neurons (forward gains of input axons to the bulb) but between bulbar neurons (lateral feedback gains within the bulb). This is a fundamentally important distinction.

     The next step was to consider how the change in synaptic strength might take place. According to the most widely held hypothesis two conditions must hold. One requirement is that both the pre- and postsynaptic neurons must be active simultaneously. This can occur if two or more receptors receive an odor on the same trial and excite two or more mitral cells that are mutually excitatory. The best candidate synapse in the bulb is axosomatic on the cell bodies of mitral cells; in electron micrographs (Fig. 7) it is reciprocal or bidirectional, having synaptic vesicles on both sides of the junction (Willey, 1973). The second requirement is for an enabling input that is delivered by a centrifugal pathway. A possible candidate is one of the aminergic projections from the brain stem into the bulb, such as the noradrenergic system that is known to be activated by appropriate unconditioned stimuli.

     The next consideration is the immense number of receptors. The exact number that is sensitive to any one odor is not known but can be estimated at one to ten million. The number that is activated on any one sniff is also unknown, but especially at low odor concentrations it is likely to be less than all, perhaps a subset of some tens to tens of thousands. Due to turbulence of air flow in the nose during a sniff, the activated subset is likely to differ with each sniff, so that over the several hundred sniffs in each training session it is likely that a major fraction of the receptors sensitive to the odor CS do receive the odor on one or more sniffs. Because the mitral cells are coactivated by the receptors in related fashion, what can emerge is a large-scale pattern of strengthened connections among a subset of mitral cells. In accordance with the concepts of Hebb (1949) I have called that subset of neutrons and their strengthened connections a template for a nerve cell assembly that corresponds to the odor CS (Freeman, 1979c).


Fig. 6. The configuration is shown for the change in shape of an averaged evoked potential (AEP) when an animal is made selectively attentive to the evoking electrical stimulus. It differs from the shape changes that accompany arousal (vigilance) and motivational changes. It implies an increase in the strength of mutual excitation (Freeman, 1979b) and not in the input synapses (i.e., feedback gain, not forward gain). From Emery and Freeman (1969).


Fig. 7. The ultrastructure is shown for an axosomatic synapse on the cell body of a bulbar projection neuron. It is likely that this is an example of a mutually excitatory synapse, which I infer is the type that undergoes modification with learning. The basis for the change is coactivation of pre- and postsynaptic cells in the presence of centrifugal enabling input. This bidirectional synapse (both presumably excitatory) seems well adapted for that role. From Willey (1973).


     According to the KII model, when a template has been established, input to any of the mitral cells within the assembly serves to active the entire assembly through their mutually excitatory connections (Fig. 8). The output of the model is a stabilized spatial image. Evidence for image stabilization in the bulb was found by comparing the mean spatial pattern of EEG amplitude over ten trials with the spatial pattern of the standard deviation of EEG amplitude. Commonly the maxima of variation in EEG were found to occur eccentrically to the maxima of amplitude, which was not to be expected for random variables (Freeman and Schneider, 1982). This property was also simulated with the KII model (Freeman, 1979c).


Fig. 8. An example is shown of simulated bursts from the KII model. A. A template was formed by coactivating three input lines and increasing by a small amount the gains of the mutually excitatory synapses in proportion to the correlation coefficients of their activity. B. A test input was given to one element. The spatial pattern conformed to the template of strengthened connections and not to the input, providing an example of the filling in of missing details. From Freeman (1979c).



     The relative invariance of the EEG spatial pattern gives an essential clue to the behavioral significance of the neural mechanism of the bulb. First, the synaptic template can provide for an active filter that selectively enhances the sensitivity of the bulb to expected odors without impairing its responsiveness to unattended odors. The suppression of responses to unwanted odors can also occur but by other neural mechanisms that effect habituation (Freeman, 1979b; 1979c). By these means the bulb can abstract selected input. Second, the bulbar activity that is governed by the template can be regarded as a generalization over particular combinations of receptor input, such that there is a characteristic output for every presentation of a given odor CS, whether or not the particular combinations occurred during the training period (Fig. 8). In psychological terms the template can provide for an invariant response over a class of equivalent stimuli. Third, the spatial pattern of bulbar activity can serve to represent an expected CS even if the particular odor is not present. For these reasons I have called the activity pattern the manifestation of an olfactory "search image" for the odor CS (Freeman, 1981), following the practice of ethologists who have established behavioral evidence for the existence of search images in numerous animal species (Atema et al., 1980).

     The search image may be said to dominate bulbar output but not to the exclusion of other output relating to other odors. For example, an unexpected odor previously given significance leads to the orienting response and an activation of a dormant template. A persisting novel odor if reinforced also induces the orienting response and formation of a new template. It is in this condition wherein the novel input is still "undefined" for the subject that the EEG odor-no odor differences are detected. One may indeed ask how far into the nervous system raw sensory input can pass; the answer is not yet known.

     If this line of reasoning is correct, then these EEG contour plots reveal our first glimpse of the physical aspect of a mental event. Even though we cannot know whether rabbits introspect, I have elected here to call it a "mental image" partly because I believe that neural and mental images are two sides of the same coin, and partly because I want to attract the attention of cognitive psychologists, who believe that such "images" or "representations" must exist, and who now vigorously debate their nature (Block, 1981). All agree that they are products of the brain and that they form under the physical and chemical constraints that govern the brain. This primitive example allows us to speculate about physical properties of brain images.

      Clearly the event under observation is analog as opposed to digital. It consists in the cooperative activity of indefinitely large numbers of neurons and their continuously variable dendritic currents, axonal pulse densities, and concentrations of chemical transmitters and modulators. Its local manifestation occupies several cubic millimeters of tissue without sharp boundaries. A proper language for description is by integrodifferential equations of classical physics and not Boolean algebra as applied to logical switching networks.


     Yet it is not an analog of that which it represents. An odorant and the neurons of the receptor layer and bulb are made of the same substance that obeys the same laws of physics and chemistry, but there is no other way in which the physicochemical properties of the odorous molecules resemble the neural activity that follows reception.

     Is it a picture or a description? During normal respiration there is a sequence of oscillatory bursts in the two surface dimensions of the bulb; in this respect the image more closely resembles the frames in a motion picture than it does a one-dimensional tape recording of a spoken sentence. Yet it is not a picture, which is a static distribution of pigments on a surface, nor does it represent literally the olfactory ambience at a given time. The pathways for lateral interaction within the bulb and cortex that shape the patterns of cooperative activity within them and the resonance properties between them suggest that this event is not denotative and descriptional (determined by forward gains) but is connotative and associational (determined by feedback gains). The traces of an odor that are stored in synaptic memory (the engram) are not unitary, serially ordered, or episodic; they are the cumulative record of the covariances that were transiently induced among central neurons by receptor input, during periods of learning in which a certain category of input was received.

     The olfactory burst is an ephemeral activity for which the pattern is subject to continual variation on successive inspirations, but it is not epiphenomenal. It is a dynamic process that draws its shape from the olfactory environment, from the residues of past environments embedded in the synaptic connections, and from the chemical state of the bulb controlled by centrifugal input from the brain. It is an operator that sweeps up neurons by the millions into coherent and cooperative activity, thereby creating information. For the subject it is the information, because it shapes the output of the bulb. It provides a "form" for Merleau-Ponty's (1961) "structure of behavior."

     Perhaps the strongest objection to use of the concept of mental image is an implicit requirement for a reader or viewer of the image in the brain. The physiological properties of the image suggest that there is no such requirement. By extrapolating the interactive properties of the bulb and cortex to other parts of the limbic system, such as the amygdalostriatum and hippocampus that have similar structures and dynamics, we can conceive of the activation or creation of images of expectant motor activity (Freeman, 1981). Already there is substantial electrographic evidence for the coordinated and cooperative activity of these several structures at common frequencies during the performance of conditioned responses (Bressler, 1982; Macrides et al., 1982). There is higher function in the neocortex as well, to be sure, but there is not infinite regression.



     It should not be lost to you that the search-image hypothesis is itself a search image. EEG waves do not carry labels to identify them as images; that interpretation stems from the context of the anatomy and physiology of the bulb, the history of behavioral testing, and the mathematical model that describes the neural dynamics, in which the key feature is the sigmoid curve. It postulates that "information processing" is a highly interactive neural process that assigns incoming information to preexisting categories by abstraction and generalization. It comes close to restating contemporary psychological views on the process of perception, but it emphasizes the creative aspect of that process. Lest this seem unexpected or ironic, coming as it does through high technology, let me add a few words on the origin of this hypothesis.

     The image I would like to invoke is that with each breath the rabbit draws it creates a new world view, and if rabbits do that then surely so do we. While we can properly regard ourselves as biological mechanisms, we are not reflex automata. The hypothesis I have sketched departs fundamentally from the past three centuries of stimulus-response determinism and its key descriptor, the reflex, which was explicitly derived by Prochaska (1784) as a metaphor of reflection in Newtonian optics. Its origin lies in a quiet revolution that has been transmuting several fields of natural science during the past two decades. New forms of mathematics and new technologies for solving equations and handling immense quantities of data have made it possible to represent and thereby comprehend very large and complex systems. In constructing the search-image hypothesis I have drawn on the metaphysical foundations and mathematical formalisms of certain aspects of electrical engineering and theoretical chemistry, particularly the work of Ilya Prigogine (1980), and adapted them to the nervous system.

     According to this perspective, when microscopic particles such as atoms, cells, insects, or neurons interact in large numbers, a macroscopic entity forms. When a source of energy and a sink for entropy are provided, then space-time patterns of energy exchange emerge within the entity. These patterns grow and evolve through sequential states in the direction of ever-increasing complexity. The key new insight is that material systems in their natural settings tend increasingly to order and not disorder. The law of entropy is not repealed but it is disenthroned. Order emerges from chaos without external prescription in a manner that calls for the term "self-ordering" or "self-organizing." This property holds for superheated plasmas in nuclear reactors; for atmospheric storms; for diffussion-coupled chemical reactions; for the dividing cells of embryos; for the behavior of social insects; and par excellence for the myriads of neurons comprising the nervous system. The brain creates information because it is a self organizing physicochemical machine. This means that information given to a rabbit by an observer in the form of a warning odor may be terminated or be overridden and be recreated in the bulb in a new form that is unique to the observed. It also means that the information carried by sound waves from my voice has terminated in your cochlear nuclei, and that each of you has recreated the information in a new form that depends largely on your prior understanding. It becomes a search image that emerges with the question "What actually did he say?" I close with a metaphor, which is a mental image. When Martin Luther King said "I have a dream," he created an image in the minds of millions. No one asked him what he saw; they knew. It impelled them to actions ranging from creative social change to assassination. That is the nature of the mental image.



Atema, J., Holland, K., and Ikehara, W. F. (1980). Olfactory responses of yellowfin tuna {Thunnus albacares) to prey odors: Chemical search image. J. Chem. Ecol. 6: 457-465.

Block, N. (ed.) (1981). Imagery, MIT Press, Cambridge, Mass.

Bressler, S. L. (1982). Spatiotemporal analysis of olfactory signal processing with behavioral conditioning. Ph. D. dissertation in Physiology, University of California, Berkeley.

Davis, G. W., and Freeman, W. J. (1982). On-line detection of respiratory events applied to behavioral conditioning in rabbits. IEEE Trans. Biomed. Eng. 29: 453-469.

Emery, I. P., and Freeman, W. J. (1969). Pattern analysis of cortical evoked potential parameters during attention changes. Physiol. Behav. 4: 69-77.

Freeman, W. J. (1975). Mass Action in the Nervous System, Academic Press, New York. Freeman, W. J. (1978). Spatial properties of an EEG event in the olfactory bulb and cortex. Electroencephalog Clin. Neurophysiol. 44: 586-605.

Freeman, W. J. (1979a). Nonlinear gain mediating cortical stimulus-response relations. Biol. Cybern. 33: 237-247.

Freeman, W. J. (1979b). Nonlinear dynamics of paleocortex manifested in the EEG. Biol. Cybern. 35: 21-37.

Freeman, W. J. (1979c). EEG analysis gives model of neuronal template-matching mechanisms for sensory search with olfactory bulb. Biol. Cybern. 35: 221-234.

Freeman, W. J. (1981). A physiological hypothesis of perception. Persp. Biol. Med. 24: 561-592.

Freeman, W. J., and Schneider, W. (1982). Changes in spatial patterns of rabbit olfactory EEG with conditioning to odors. Psychophysiology 19: 44-56.

Hebb, D. O. (1949). Organization of Behavior, Wiley, New York.

Lancet, D., Greer, C. A., Kauer, J. S., and Shepherd, G. M. (1982). Mapping of odor-related neuronal activity in the olfactory bulb by high-resolution 2-deoxyglucose autoradiography. Proc. Natl. Acad. Sci. (U.S.) 79: 670-674.

Macrides, F., Eichenbaum, H. B., and Forbes, W. B. (1982). Temporal relationship between sniffing and the limbic theta rhythm during odor discrimination reversal learning. J. Neurosci. 2: 1705-1717.

Prigogine, 1. (1980). From being to becoming: Time and complexity in the physical sciences W. H. Freeman, San Francisco.

Prochaska, J. (1784). Adnotationum Academicarum, Wolfgang Gerle (trans.), Prague.

Laycock, T. (1851). The Principles of Physiology by John Augustus Unzer; and a Dissertation on the Functions of the Nervous System by George Prochaska. London, Sydenham Society.

Willey, T. J. (1973). The ultrastructure of the cat olfactory bulb. J. Comp. Neurol. 152: 2 1 1-23 2.



- E N D -