The ability to image the fully functioning brain offers the tantalizing hope of finally capturing one of life's amazing moments—the instant we learn something. Since the time of Aristotle, we have wanted to know what happens when we learn—how an association is formed and where it is formed. Is it, as Descartes might have supposed, a change in the flow of vital animal spirits at the pineal? Does it even take place in the brain or, as the ancients proposed, in the heart? After all, we still talk about learning something “by heart.” Actually, it has become clear from a large number of experiments using a range of traditional biological techniques, including lesions and neural recordings, that associations are indeed formed in specific areas of the brain (Thompson, 1986) and involve changes in the membrane properties of identifiable cells (Alkon et al., 1987; Schreurs et al., 1998). Although these studies have informed us about the necessity and sufficiency of particular structures and cell types involved in learning, we still do not know how an intact, integrated, fully functioning brain forms an association.
To begin to answer this question, we will review some of the basic behavioral issues involved in how associations are formed, as well as the methods and the model systems that have been used to study associations. We will then review what is known about the formation of associations in two model systems from standard anatomic and physiologic studies to lay the groundwork for describing the advances that imaging studies have helped make in our understanding of how learning and memory take place (Table 1).
Table 1. List of abbreviations related to the study of learning and memory
Excitatory postsynaptic potential
Long-term synaptic depression
Magnetic resonance imaging
Nuclear magnetic resonance
Positron emission tomography
Protein kinase C
Regional cerebral blood flow
Reverse transcription-polymerase chain reaction
Associative learning is a relatively permanent change in behavior resulting from the temporal conjunction of two or more events. Although there are important caveats such as “silent” learning, in which learning occurs even when observable behavior is blocked (Mackintosh, 1983; Krupa et al., 1996), the current definition of associative learning has heuristic value. Obviously, not every two events that occur together are associated. Aristotle proposed that previous contiguity, as well as similarity and contrast, play a role in which events would be recalled together. Later, the British Empiricists postulated the Laws of Association as a means of formally stating which events could become associated. The Law of Contiguity stated that an associative connection between two events would only be formed if they occurred together in time and space. The Secondary Laws of Association addressed the frequency with which events occurred in contiguity, the duration of the events, their intensities, the number of other associations in which the two events had already been involved, the similarity of the association to other associations, and the abilities, emotional state, and bodily state of the person experiencing the events (Gormezano and Kehoe, 1981).
Ebbinghaus, who reported his own ability to learn and then relearn pairs of nonsense syllables, conducted the first experiments on the laws of association in 1885. However, the study of associations received its first major impetus in Russia where, in the early 1900s, classical conditioning was proposed as the prototype of associative learning (Pavlov, 1927).
With the growing interest in classical conditioning as a means of studying associative learning, a need arose for the definition and characterization of classical conditioning, and a need for the specification of appropriate control procedures (Gormezano and Kehoe, 1975). One such characterization includes (1) the presentation of an unconditioned stimulus (US) that reliably elicits an unconditioned response (UR); (2) the use of a conditioned stimulus (CS) that has been shown by test to produce a response that initially does not resemble the UR; (3) the repeated presentation of the CS and US with a specified order and temporal interval; and (4) the emergence of a new response to the CS, the conditioned response (CR), which is similar to the UR (see Figure 1).
A broader definition of classical conditioning has emerged in the past 20 years in which it has been described as the study of the relations among stimuli in the environment.
A broader definition of classical conditioning has emerged in the past 20 years in which it has been described as the study of the relations among stimuli in the environment (Rescorla, 1988). In this broader context, the CS is said to signal the US during pairings and the question that is asked is whether exposure to the relation between the CS and US modifies the behavior of the organism in a detectable way. For example, if an organism shows an increased response to the CS as a result of being exposed to a relationship between the CS and US, some have suggested that an association has formed between the two events. However, as we shall see in a discussion of control procedures, there are several nonassociative contributors to responding that may also produce increases in responding to the CS.
Although repeated pairings of a CS and US may lead to the emerging of a CR, the occurrence of a response to the CS may result from nonassociative, as well as associative, processes (Gormezano and Kehoe, 1975). The first nonassociative factor that may influence responding is the baseline level of activity that occurs in most response systems. The second factor is the occurrence of a response to the CS in untrained animals. For example, a bright light can elicit a reflexive blink in many species, including humans (Grant, 1943; Grant, 1944). A bright light may also change heart rate, respiration, and the electrodermal response. If the eye blink is the response of interest, a reflexive or unconditioned blink to the light has been termed an “alpha” response (see the top trace in Figure 1). A change in the reflexive blink as a result of CS-US pairings is called “alpha conditioning” (Grant and Adams, 1944). The third potential nonassociative factor is the sensitizing effect that the US may have on baseline responding or CS-elicited responses. To assess the contribution of nonassociative processes to responding, researchers have adopted control procedures incorporating US-alone, CS-alone, and explicitly unpaired presentations of the CS and US.
A broader approach to control procedures, epitomized by the “truly random” control, is based on manipulating the degree of the relationship between the CS and US (Rescorla, 1988). In the explicitly unpaired control procedure, the CS and US never occur together, so there is a negative relationship between the CS and US. In fact, the CS is thought of as perfect predictor of the absence of the US. On the other hand, the truly random control procedure consists of independently programmed occurrences of the CS and US that are presented in an attempt to ensure that there is no consistent relationship between them. In the truly random control, there is an equal probability of US occurrence in both the presence and absence of the CS (Rescorla, 1988). Experiments on the truly random control procedure have shown that some conditioning does occur because of fortuitous CS-US pairings that happen when there is an equal probability of US occurrence in the presence and in the absence of the CS (Kremer and Kamin, 1971; Benedict and Ayers, 1972; Ayers et al., 1975; Rescorla, 2000).
The bottom line is that we can study the formation of an association in the laboratory by using classical conditioning paradigms and their appropriate controls. In the most basic procedure—delay conditioning—a subject is repeatedly presented with a pair of stimuli, such as a tone and air puff. On trial n, the subject responds with a blink to the puff. On trial n+1, the subject is again presented with the tone and puff, but this time responds with an anticipatory or conditioned blink to the tone, as well as blinking to the puff. The fascinating question is what has occurred in the subject to produce this change in behavior? Although the stimuli are the same, a new response has now emerged. This new response shows us that an association has been formed between the two events. In short, learning has taken place.
A large number of different animal paradigms have been used to study the neural substrates of learning and memory (for review see Schreurs, 1989). We will briefly review two model systems, one invertebrate and the other vertebrate, that have been used to image learning and memory.
Classical conditioning in the marine snail Hermissenda has been demonstrated in a procedure in which 3 seconds of light (CS) are presented together with 2 seconds of rotation (US) (Alkon, 1983). Usually, rotation of the animal elicits clinging with the body musculature contracted and the foot gripping the bottom of the enclosure (see Figure 2). Presentations of the light alone to animals placed in the dark during the light portion of their diurnal cycle produces the phototropic response of movement toward the light source. As a result of light-rotation pairings, classically conditioned animals have a significantly longer latency to move toward the light than control animals that received rotation alone, random but separate presentation of light and rotation, light alone, or strictly alternating presentations of light and rotation. The learned increase in latency to move toward light is a function of the number of CS-US pairings and CS-US interval. The conditioned clinging can be extinguished and is specific to the light CS; it displays savings and is sensitive to CS/US contingencies.
In addition to learning-specific changes in movement to light, trained animals also show a new response to light—foot contraction—which emerges as a function of training. Observations of Hermissenda in the dark revealed a 15–20% shortening in the length of the foot (the single organ of locomotion) in response to rotation (Figure 2). It was also noted that there was a small lengthening of the foot in response to light before training. After pairings of light and rotation, the foot contraction elicited as an unconditioned response to rotation emerged as a new response to light in all of the trained animals. The new response of foot contraction to the light did not occur in untrained animals nor did it occur in animals that received random presentations of light and rotation (Lederhendler et al., 1986). Simultaneous measurement of both locomotion and foot contraction revealed that there is a rapid decrease in locomotion and a gradual increase in foot contraction as a function of training (Matzel et al., 1990a). Finally, foot contraction was also shown to be a function of the number of CS-US pairings and of the interval between the CS and US (Matzel et al., 1990b).
The sensory inputs involved in Hermissenda associative learning include cells of the visual system and a primitive vestibular system (Alkon, 1983). The motor output neurons involved in classical conditioning of foot contraction include motoneurons controlling the muscles of the foot and turning of the animal toward a light source. The visual system consists of two relatively primitive eyes, each containing a lens, pigment cells, and five photoreceptors (three type B cells and two type A cells). The vestibular system consists of two statocysts that contain 12 hair cells and a cluster of crystals called stataconia, which brush up against the hair cells when there are changes in the direction of gravitational force. In Hermissenda, there is convergence between the visual and vestibular pathways at the level of the sensory cells themselves (Figure 3A). There is also convergence of the sensory inputs on interneurons that, in turn, connect to the motor outputs. A detailed analysis of the networks within and between the visual and vestibular pathways has been possible (Figure 3B). Knowledge of this neural organization has allowed a step-by-step tracking of training-elicited signals as they flow through the identified pathways (Alkon, 1987).
Pairings of light and rotation lead to an increase in the excitability of the type B photoreceptor and a decrease in the excitability of the type A receptor. The changes in excitability are the net effect of interactions of several inputs to the sensory circuitry of Hermissenda. First, light excites the B photoreceptors, and rotation excites the hair cells, which, in turn, inhibit the B photoreceptors. When rotation stops, hair-cell activity is reduced even below unstimulated levels so that the B photoreceptor is released from hair-cell inhibition. Second, when light and rotation are paired, excitation of the B photoreceptor allows the inhibitory signals of the hair cell on the B photoreceptor to be “shunted,” which lessens the effect of rotation-induced inhibition on that B photoreceptor. Third, when light and rotation are stopped, there is an increase in the excitatory feedback from the second-order visual cells of the optic ganglion onto the B photoreceptor. Fourth, a transmitter released by the hair cells (gamma-aminobutyric acid, GABA) onto the B photoreceptor causes prolonged depolarization and potassium current reduction. These effects of GABA are enhanced by B photoreceptor depolarization, which is maximized when light is paired with rotation during training. The cumulative effect of these four sources of increased excitation is an increase in the overall excitability of the B photoreceptor so that its response to light is enhanced and prolonged. One of the results of increased B photoreceptor excitability is the inhibition of a chain of neurons responsible for motor neuron impulses that drive the muscles, which causes a turning of the foot. Another result is an increased excitation of interneurons driven by hair cells that then excites motoneurons controlling foot contraction.
The pairing-specific excitability of the type B photoreceptor has been shown to result from an increase in intracellular calcium that interacts with an elevation of diacylglycerol (a membrane-bound lipid), and probably arachidonic acid, to induce movement of the calcium-sensitive enzyme protein kinase C (PKC) from the cell cytoplasm to the cell membrane. In the cell cytoplasm, PKC increases potassium ion flow, which, in turn, decreases membrane excitability. In the cell membrane, PKC reduces potassium ion flow, which, in turn, increases membrane excitability. Agents that block the movement of PKC from the cytoplasm to the membrane also block the decrease in potassium ion flow. Translocation of PKC to the inner surface of the neuronal membrane allows it to interact directly with protein targets for phosphorylation (Alkon, 1989). In Hermissenda, and later in the rabbit hippocampus (see below), a particular calcium- and GTP-binding protein known as calexcitin was shown to undergo phosphorylation only during classical conditioning procedures. Recently, calexcitin has been cloned and sequenced completely. This protein has been shown to be a high affinity substrate for the alpha-isozyme of PKC. The alpha-isozyme of PKC, although not limited to neuronal tissue, is by far the most abundant of the PKC isozymes in the brain.
Calexcitin serves as a signaling molecule that amplifies calcium elevation in response to learning-associated synaptic transmitters that initiate second messengers such as diacylglycerol, arachidonic acid, and calcium itself.
PKC-mediated phosphorylation of calexcitin also causes the latter to translocate to neuronal membranes of three principal types: the outer wall membrane, the membrane of the endoplasmic reticulum (ER), and the nuclear membrane. These translocation targets are consistent with the recently revealed pleiotropic functions of calexcitin, which are consistent with and apparently required for synaptic modification during learning (Alkon et al., 1998). At the outer wall membrane, calexcitin directly inactivates voltage-dependent potassium currents, just as it occurs during classical conditioning to cause increased excitability (and thus, increased synaptic weight). At the nuclear membrane, calexcitin has been found to increase turnover of mRNA, presumably for several learning related proteins. Other studies had demonstrated previously that there is a close correlation of Hermissenda classical conditioning with increased turnover of mRNA for a variety of protein species.
At the ER membrane, calexcitin has now been shown to bind the ryanodine receptor (RYR) with high affinity (Nelson et al., 2001). Direct measurements of efflux at isolated ER membrane vesicles, in fact, demonstrated that calexcitin is a specific endogenous ligand necessary for calcium-mediated calcium release by means of the RYR. Therefore, calexcitin serves as a signaling molecule that amplifies calcium elevation in response to learning-associated synaptic transmitters that initiate second messengers such as diacylglycerol, arachidonic acid, and calcium itself. The crucial role of the RYR was also recently confirmed with gene fingerprinting technology that revealed prolonged learning-specific enhancement of RYR mRNA expression for many hours after rat spatial maze learning. Enhanced RYR expression was further confirmed by quantitative RT-PCR, Northern blots, and other molecular biologic techniques such as in situ hybridization.
Rabbit Eye Blink/Nictitating Membrane Response
Classical conditioning of the rabbit nictitating membrane response (NMR) and eye blink was first reported when Gormezano and his colleagues paired a tone CS with a corneal air puff US (Gormezano et al., 1962; Gormezano, 1966). Usually, corneal air puff elicits extension of the nictitating membrane and closure of the outer eyelids (eye blink) whereas the tone does not elicit that response (Figure 4). Animals given paired CS-US presentations show the emergence of an NMR to tone and a progressive increase in the frequency of conditioned nictitating membrane extension across days of training to a level of 95%. In marked contrast, animals given CS-alone, US-alone, or explicitly unpaired CS-alone and US-alone presentations never exceeded a frequency of 6% membrane extension on any single day and averaged a level not appreciably higher than the base rate of 2–3%. Extensive behavioral experiments have shown that the acquisition of the conditioned NMR displays all the hallmarks of classical conditioning, including sensitivity to the number of CS-US pairings, CS-US interval, CS and US intensity, CS specificity, and savings.
The sensory inputs of the rabbit nictitating membrane response include the brain stem auditory pathways and corneal inputs that travel by means of the spinal trigeminal nucleus to the inferior olive and cerebellum, and on to the motor output of the accessory abducens nucleus that controls eyeball retraction and the resultant sweep of the nictitating membrane. In addition to these primary pathways, CS and US information also travel to many other parts of the brain, including the hippocampus, cerebellum, and cortex. A great deal of lesion and recording research, initiated in large part by Thompson and his colleagues, has revealed the important role played by the hippocampus and cerebellum (Thompson, 1986). For example, neural activity that mimics and precedes conditioned nictitating membrane responding has been recorded in the hippocampus. Electrodes placed in the region of the CA1 pyramidal cells of the hippocampus show that the CA1 cells respond to a tone CS at a higher frequency and shorter latency as a consequence of classical conditioning of the NMR, rather than as a consequence of unpaired stimulus presentations.
Intracellular recording of CA1 cells in a slice of hippocampus obtained from classically conditioned rabbits revealed a reduction of the flow in potassium ions through the cell membrane in much the same way as in the Hermissenda Type B photoreceptor. In fact, voltage-clamp recordings in CA1 cells showed that potassium-ion currents active in the presence of calcium were modified as a function of classical conditioning (Sanchez-Andres and Alkon, 1991). This current is similar to one of those implicated in the associative learning demonstrated by Hermissenda.
Neural recording in the cerebellum during classical conditioning of the rabbit NMR has shown both increased and decreased extracellular activity in the dentate/interpositus deep nuclei and cortex correlated with the CS, US, CR, and UR (for review see Thompson and Kim, 1996). Some studies recording Purkinje cell activity in intact behaving animals have shown decreases in activity in the CS period as a result of training (e.g., Thompson, 1986). Several other studies have shown both increases and decreases in Purkinje cell activity (e.g., Gould and Steinmetz, 1996). Lesions of the deep nuclei of the cerebellum have been shown to abolish CRs in trained rabbits and prevent the acquisition of CRs in naive rabbits. Lesions of the cerebellar cortex, particularly lobule HVI, have severely impaired CRs, and bilateral lesions of HVI have abolished CRs.
Several years ago, we adopted the cerebellar slice preparation to identify the cellular correlates of these conditioning-specific changes in Purkinje cell activity (Schreurs et al., 1991, 1997a, 1998). In a typical experiment, rabbits are given sessions of paired or explicitly unpaired presentations of a tone and electrical stimulation around the right eye. Twenty-four hours later, slices are prepared of lobule HVI from the right side of the cerebellum and Purkinje cell electrophysiological properties are measured, including membrane excitability (measured as threshold for dendritic spikes, Schreurs et al., 1991, 1997a, 1998), synaptic excitability (measured as current required to elicit an excitatory postsynaptic potential [EPSP], Schreurs et al., 1997a), and potassium channel function (measured as the effects of potassium channels antagonists, Schreurs et al., 1998).
Our experiments show that membrane excitability was higher in Purkinje cell dendrites in lobule HVI of rabbits given paired stimulus presentations than of rabbits given unpaired stimulus presentations Schreurs et al. (1991, 1997a, 1998). This excitability was indexed, in part, by a lower minimum current required to elicit dendritic calcium spikes in cells from paired animals than in cells from unpaired animals. A second index of a change in excitability was a decrease in the size of a potassium channel-mediated transient membrane hyperpolarization in cells from paired rabbits (Schreurs et al., 1998). A third index of excitability was a decrease in the threshold current required to elicit EPSPs and Purkinje cell spikes, which was found to be lower in cells from paired animals than in cells from unpaired control animals (Schreurs et al., 1997a).
The increase in Purkinje cell excitability was found to be highly correlated with the acquisition of CRs and was preserved for as long as 1 month after conditioning (Schreurs et al., 1998). In addition, the ability to induce long-term depression in Purkinje cells of lobule HVI was occluded by classical conditioning (Schreurs et al., 1997a). These data suggest that the conditioning-induced increase in Purkinje cell excitability prevents long-term synaptic depression (LTD) and that cerebellar LTD may not be the mechanism underlying conditioning of the rabbit NMR. However, it should be noted that studies using knockout and mutant mouse models suggest a correlation between cerebellar cortical LTD and eye blink conditioning (see Kim and Thompson, 1997). Finally, pharmacologic experiments implicating a specific potassium channel (Schreurs et al., 1997a) suggest that PKC-calexcitin mediated changes in potassium channels of the type found in Hermissenda B photoreceptors and that rabbit hippocampal CA1 cells may be the mechanism responsible for the observed increase in Purkinje cell dendritic excitability.
Experiments designed to provide images of learning and memory can be divided into two distinct categories. The first category includes animal studies that, with few exceptions, have focused on the functional labeling of fixed tissue. The second category includes human studies that have focused on in vivo functional imaging techniques. As an important methodologic note, it should be emphasized that all of the studies reviewed in the following sections contain critical control groups. These control groups are critical because they help to identify learning-specific functional changes and not simply functional changes that occur over the course of the experiment. Despite the use of adequate controls in many imaging studies, the search for localized function of complex cognitive processes has prompted some to speak of a “new phrenology” (Uttal, 2001).
Animal experiments have focused, in large part, on obtaining a steady state picture of the brain at some point after learning has commenced. These studies almost invariably compare an animal that has been trained with a control. One group of animal imaging studies has focused on mapping electrical activity that is indexed by processes including oxidative energy metabolism, as described by Sokoloff and colleagues (Sokoloff et al., 1977), and expression of the immediate early gene c-fos (Irwin et al., 1992; Carrive et al., 1997; Gruart et al., 2000). A second group of studies has been based on imaging learning-specific cellular processes that have been identified in model systems (Olds et al., 1989, 1990; Scharenberg et al., 1991; McPhie et al., 1993).
Glucose is the primary source of energy in the brain. The majority of glucose is used by the mitochondria of the cell to provide ATP to the sodium/potassium pump. However, glucose is metabolized rapidly to CO2 and H2O, which makes it difficult to image. To image glucose utilization, Sokoloff and colleagues (1977) radiolabeled an analog of glucose, 2-deoxy-D-glucose (2-DG), which is a competitive substrate for blood-brain transport and phosphorylation but is not metabolized to CO2 and H2O and is retained in the tissue (Sokoloff, 1992). The metabolism of glucose analogs such as 2-DG can be used to map membrane electrical activity changes maintained by the sodium/potassium pump (Gonzalez-Lima, 1992). Gonzalez-Lima and colleagues have used 2-DG and related analogs such as fluorodeoxyglucose (FDG) to image by autoradiography evoked neural activity to auditory conditioned stimuli following classical conditioning procedures in rodents (Gonzalez-Lima and Scheich, 1984, 1986; Gonzalez-Lima, 1992; McIntosh and Gonzalez-Lima, 1995, 1998).
The goal of these experiments was to visualize neural activity changes by obtaining a learning-specific change in a quantitative functional index of neural activity (Gonzalez-Lima, 1992). The main finding was an enhanced metabolic response produced by pairings of an auditory CS with aversive midbrain reticular formation stimulation as a US. The response was learning-specific, because it was significantly higher than levels of metabolic activity in several control groups, including subjects that received CS-alone, US-alone, pseudo-random, and backward stimulus presentations. Learning-specific increases in FDG uptake were found in the dorsal and ventral cochlear nuclei, the first two nuclear stations of the central auditory system (Gonzalez-Lima and Scheich, 1984), and up into the auditory cortex (Gonzalez-Lima and Scheich, 1986).
Harvey and colleagues (1988) also noted the involvement of the primary auditory pathway in learning when they used uptake of deoxyglucose (2-DG) in tissue samples to determine changes in rabbit brain during classical conditioning of the rabbit NMR. They compared paired and explicitly unpaired rabbits after 6 days of tone and air puff presentations, and found a significant pairing-specific increase in uptake of 2-DG in tissue taken from the dorsal cochlear nucleus.
External signals may activate genes that encode transcription factors that modify other genes. The first genes activated by external signals are those that do not require de novo synthesis of proteins and are known as immediate early genes. The most widely studied of these genes is c-fos. The proto-oncogene c-fos is activated in the brain by a large number of stimuli including seizure activity, stress, and noxious stimulation and has been used to map brain metabolism under different physiological conditions (Herrera and Robertson, 1996). The presence of c-fos-like activity can be detected with immunohistochemistry by using poly- or monoclonal antibodies to various fragments of c-Fos protein and subsequent visualization. Several groups have examined either c-fos or Fos expression following classical conditioning of the rabbit NMR by using an aversive US such as air puff or periorbital electrical stimulation (Irwin et al., 1992; Carrive et al., 1997; Gruart et al., 2000). In a relatively early study, Irwin et al. (1992) reported a conditioning-specific increase of c-fos expression in the raphe nucleus and a conditioning-specific decrease of expression in the trigeminal nucleus. Carrive et al. (1997) reported a conditioning-specific increase in Fos-like immunoreactivity in the locus coeruleus after classical conditioning of the rabbit NMR. Finally, Gruart et al. (2000) detailed considerable conditioning-specific expression of Fos in the hippocampus and the occipital, parietal, piriform and temporal cortices.
As we have already noted, PKC is an enzyme that seems to play an important role in learning and memory. It is found in Hermissenda and is highly enriched in vertebrate brain tissue, including the hippocampus and cerebellum. PKC has been shown to play a critical role in neural plasticity, learning, and memory (Akers et al., 1986; Worley et al., 1986; Bank et al., 1988; Saito et al., 1988; McPhie et al., 1993; Sunayashiki-Kusuzaki et al., 1993). In several studies, Alkon and colleagues had shown a long-term translocation of PKC from the cytosol to the membrane after classical conditioning of the rabbit NMR (Bank et al., 1988; Coulter et al., 1989). Imaging studies have now shown membrane-specific changes in PKC binding after classical conditioning in Hermissenda and the rabbit, as well as in rat olfactory discrimination and water maze learning (Olds et al., 1990).
Measurements of PKC after classical conditioning of the rabbit NMR indicate that there is an increase in membrane-associated PKC near the rabbit CA1 cell bodies even 1 day after all training has been completed (Olds et al., 1989). Three days after training (i.e., well into the period of memory retention), maximal PKC labeling moved from the cell bodies to the region of the CA1 dendrites (Figure 5). Interestingly, the movement of PKC in CA1 cells can be artificially induced by the drug phorbol ester, which also causes the same potassium ion flow reduction that takes place during conditioning (Alkon, 1989). Translocation of PKC to the CA1 membrane by phorbol ester also causes enhanced summation of EPSPs elicited by activation of presynaptic fibers known as Schaeffer collaterals. This same enhanced EPSP summation was demonstrated to occur only in rabbits previously trained with a classical conditioning procedure. These remarkable biophysical and molecular parallels between Hermissenda and rabbit classical conditioning also included a learning-specific increase in phosphorylation of calexcitin and a calexcitin-like substrate in both of these very different species. Thus, conservation of a molecular sequence for memory storage was implicated by parallel cellular events in molluscan and mammalian memory paradigms.
Conservation of a molecular sequence for memory storage was implicated by parallel cellular events in molluscan and mammalian memory paradigms.
This conservation was given further support by findings of memory-specific translocation of PKC in the hippocampus and other brain structures for other learning paradigms such as cue and platform spatial maze learning and rat olfactory discrimination learning in rats. Furthermore, different laboratories using different paradigms and different methodologies (e.g., monoclonal antibodies for PKC) independently confirmed the role of PKC in associative memory (for review, see Van der Zee and Douma, 1997).
The measurement of conditioning-specific changes that take place during classical conditioning of the rabbit NMR in the cerebellum have revealed an increase in membrane-bound protein kinase C that was specific to lobule HVI of paired animals (Figure 6) but was not found in lobule HVI of unpaired or sit control animals, nor was it found to be different in the dentate/interpositus nuclei of any group (Freeman et al., 1998). Finally, steady state imaging of mRNA levels (by in situ hybridization) and protein levels (by immunohistochemical measures) have demonstrated learning-specific expression patterns of several other molecules related to PKC activation. These include Type 2 ryanodine receptor (Figure 7), insulin receptor, MAPKinase, and the non-receptor tyrosine kinase (Alkon et al., 1998).
Functional imaging in humans is based on the observation, first reported more than a century ago, that cerebral activity and blood flow are connected (Roy and Sherrington, 1890). Sokoloff first demonstrated experimentally the relationship between neural activity and blood flow when he measured changes in blood flow to visual stimulation (Sokoloff, 1961). As noted above, Sokoloff and colleagues have also shown that glucose metabolism can be used to index neural activity (Sokoloff et al., 1977). These findings culminated in the fact that clinical scanning devices can now be used to detect changes in human cerebral metabolism and perfusion (blood flow, volume, and oxygenation) that result from increases in neural activity (Cohen, 1996).
Functional imaging using PET takes advantage of the fact that radioisotopes that decay by positron emission can be introduced into the blood stream intravenously. Compounds such as H215O and 18F-deoxyglucose are synthesized, injected, and used to measure blood flow and glucose metabolism, respectively. As these radiolabeled compounds circulate, they emit positrons that collide with electrons and annihilate, resulting in the emission of two 511 keV gamma rays. The greater the blood flow that occurs, the larger the number of gamma rays that are emitted and the greater the signal that can be detected. Spatial resolution of the order of 6–8 mm is typical and is fundamentally determined by the energy of the emitted positron. Higher energy positrons travel farther from the actual site of activation before annihilating; thus, spatial resolution is limited by the chosen radiochemical. Other molecules of interest can be synthesized that allow receptor-binding assays for transmitters such as dopamine and serotonin or enzymes such as monoamine oxidase (Cherry and Phelps, 1996).
Functional imaging using MRI takes advantage of the intrinsic magnetic properties of blood and tissue to provide structural and functional information. The nucleus of the hydrogen atom (proton) in a static magnetic field responds to radio frequency pulses by emitting detectable radio waves. The precise application of magnetic fields, delivery of radio frequency pulses, and detection of the resultant signal make up some of the major demands of functional MRI (fMRI). As noted above, neural activity increases blood flow and volume. However, the brain cannot use all of the increased oxygen, making the oxygen content of venous blood increase during neural activity (Cohen, 1996). The fMRI blood oxygen level detection (BOLD) contrast fMRI method takes advantage of the imbalance between the higher MRI signal from oxygenated (oxyhemoglobin) and the lower signal from deoxygenated (deoxyhemoglobin) blood to provide a time series of images correlated with neural activity.
The advent of functional imaging as a means of studying the brain has led to an explosion of studies of human learning and memory. A relatively recent review of functional imaging studies that focused on PET and fMRI identified 275 articles that explored cognitive function (Cabeza and Nyberg, 2000). Of these, over 200 could be categorized as studies involving learning, memory, or both (see also Gabrieli, 1998). For the purposes of the current review, we will narrow our focus to functional imaging studies that have examined classical conditioning by using adequate control conditions. Classical conditioning in humans has centered on paradigms such as eyelid, jaw, and limb movement conditioning, and fear conditioning (Molchan et al., 1994; Hugdahl et al., 1995; Logan and Grafton, 1995; Blaxton et al., 1996; Timmann et al., 1996; Schreurs et al., 1997b, 2001; Hugdahl, 1998; LaBar et al., 1998; Knight et al., 1999; Buchel and Dolan, 2000; Fischer et al., 2000; Maschke et al., 2000).
The advent of functional imaging as a means of studying the brain has led to an explosion of studies of human learning and memory.
A recent bibliography indicated that almost 500 human eyelid conditioning studies had been conducted before 1986 (Gormezano, 2000). Beginning in the 1920s, these studies reached a zenith during the 1960s when over 250 studies reported the results of eyelid conditioning experiments. However, not a single study examined the neural substrates of human eyelid conditioning. This trend began to be reversed indirectly in the 1990s by studies that examined eyelid conditioning in clinical populations, including Alzheimer, Korsakoff, Parkinsonian, and cerebellar lesion patients (Daum et al., 1996; Woodruff-Pak et al., 1996; Timmann et al., 1998; Sommer et al., 1999; Sears et al., 2000; Brawn Fortier et al., 2000). The issue of neural substrates was more directly addressed by another group of studies using functional imaging to study explicitly the areas of the brain involved in human eyelid conditioning in normal subjects (Molchan et al., 1994; Logan and Grafton, 1995; Blaxton et al., 1996; Schreurs et al., 1997b, 2001; Ramnani et al., 2000).
One of the first studies to image eyelid conditioning in humans (Figure 8), showed that during classic conditioning of the human eye blink response there was a significant conditioning-specific increase in regional cerebral blood flow (rCBF) in the primary auditory cortex, and a significant decrease in rCBF in the cerebellum (Molchan et al., 1994). An interesting aspect of the Molchan et al. (1994) data was the considerable laterality in the conditioning-specific changes in blood flow. Specifically, changes in the auditory cortex were strongest on the side opposite (left side) to the air puff delivery (right eye), whereas cerebellar changes were strongest on the same side as the air puff. Interestingly, the tone was presented binaurally, and observations of human eye blink conditioning on the right eye showed that conditioned responses were markedly bilateral (Hilgard and Campbell, 1936). In addition to lateralized changes in auditory cortex and cerebellum, Molchan et al. (1994) identified several other areas, including cingulate cortex, parietal cortex, and prefrontal lobes that were involved in eye blink conditioning. These data suggest that a form of learning as apparently simple as eye blink conditioning appears to engage an extensive network of neural systems. Because the involvement of these areas was specific to the acquisition of the conditioned response, the data by Molchan et al. (1994) lend support to the notion that learning and memory may involve the interactions among several neural systems and that any region has the potential to play a role, depending on the requirements of the task (McIntosh and Gonzalez-Lima, 1998).
A second study by the same group (Schreurs et al., 1997b) reversed the side of air puff delivery and was able to replicate and extend their previous findings by showing that classical conditioning of the eye blink response produced increases in rCBF in the primary auditory cortex opposite the side of air puff delivery and decreases in rCBF in the cerebellar cortex on the same side as the air puff. Several others areas associated with aspects of learning and memory were also found to change as a function of classical conditioning of the eye blink response. In particular, there were increases in rCBF in auditory association cortex and temporo-occipital cortex and decreases in the temporal poles and inferior prefrontal lobe. Additional analysis identified that the middle prefrontal cortices and the hippocampus were also involved in the associative process. However, there were also several differences between the first and second study. First, the extent of the activations and deactivations do not appear as large in the study by Schreurs et al. (1997b) as those of Molchan et al. (1994). Second, there were several areas identified in the study by Molchan et al. (1994) (e.g., striatum, parietal, and insular cortices) that did not appear in the experiment by Schreurs et al. (1997b), and several areas identified in the latter (e.g., hippocampus, temporo-occipital cortex, temporal poles) that were not found in the former study. Some of these differences may be explained by the fact that the overall levels of conditioning were somewhat lower in the study by Schreurs et al. (1997b) (62.5% CRs) than in the experiment by Molchan et al. (1994) (73.7% CRs). However, all of the areas identified across these two studies have been found to be involved in associative processes, attentional processes, or both (e.g., inferior temporal cortex, occipitotemporal cortex), so the differences between studies may reflect the phasic involvement of different areas during different stages of response acquisition.
There are two other imaging studies of human eyelid conditioning (Logan and Grafton, 1995; Blaxton et al., 1996) that provide significant similarities to as well as some differences from Molchan et al. (1994) and Schreurs et al. (1997b). Logan and Grafton (1995) examined glucose metabolism and found increases in several regions, including cerebellum, hippocampus, striatum, temporal gyrus, and the occipitotemporal fissure. Blaxton et al. (1996) examined rCBF and found increases in striatum and hippocampus and decreases in cerebellum. Clearly, several areas, most notably the cerebellum and hippocampus, were identified to be specifically involved during eye blink conditioning in all of the imaging studies. Areas such as the striatum, temporal gyrus, and temporo-occipital cortex were found to be important in at least two of the studies. Differences in imaging methods (e.g., glucose metabolism vs. rCBF, different field of views, the Logan and Grafton  study positioned the PET camera more ventrally), behavioral paradigms (massed vs. continuous training), and conditioning levels may have accounted for some of the between-study differences in brain areas and differences in the direction of changes observed in the imaging of human eye blink conditioning. These factors may also combine with sampling different time windows during the acquisition of the conditioned response. It is possible that certain areas are recruited at different times during acquisition and may not be identified depending on what portion of the acquisition curve is sampled. Taken together, the PET studies of eyelid conditioning show that several different areas participate in conditioning of the human eyelid response in an intact, normal brain.
A single, whole brain, event-related fMRI study of eyelid conditioning (Ramnani et al., 2000) has reported conditioning-specific hemodynamic response changes after differential classical conditioning. Differential conditioning differs from the standard delay classical conditioning used in the PET studies mentioned above in that a second stimulus such as a tone of different frequency, designated CS-, is interspersed among paired presentations of the usual tone (CS+) and air puff. In a further complexity, the study by Ramnani et al. (2000) omitted some of the air puff presentations on CS+ trials to assess the effects of expectancy and error correction. Conditioning-specific hemodynamic response changes were identified in the ipsilateral cerebellum, contralateral motor cortex, and hippocampus. Error-related hemodynamic changes were identified in the contralateral cerebellum and somatosensory cortex (Ramnani et al., 2000).
There are two published studies of functional imaging during classical conditioning of a flexion response (Timmann et al., 1996; Maschke et al., 2000). Timmann et al. (1996) studied leg flexion in response to presentations of electrical stimulation of the plantar nerve that was preceded by a tone CS. They showed that PET activation in the ipsilateral cerebellum, bilateral hippocampus, and bilateral frontal regions, was correlated with the level of conditioning (Timmann et al., 1996). Although Timmann et al. (1996) conducted unpaired presentations of the CS and US before the paired presentations, no explicit comparisons in PET activation were conducted between the two phases. In the second flexion study, Maschke et al. (2000) studied jaw opening in response to electrical stimulation of the mouth that was preceded by a tone CS. They found that a third of the subjects showed CRs and that these subjects had significant PET activation in the cerebellum, temporal lobe, frontal lobe, and thalamus (Maschke et al., 2000). The authors report that the remaining five subjects did not show CRs nor show any PET activation in the cerebellum. It should be noted that the jaw-opening reflex habituates quickly and that the response measure was a decrease in muscle activity that may be difficult to detect because subjects were requested to preactivate their jaw-closing muscles (Maschke et al., 2000).
The majority of recent functional imaging studies of classical conditioning have focused on “fear conditioning” in which a neutral auditory or visual stimulus is paired with an aversive stimulus such as a loud noise or shock (Cahill et al., 1995; Hugdahl et al., 1995; Buchel et al., 1998, 1999; Hugdahl, 1998; LaBar et al., 1998; Knight et al., 1999; Buchel and Dolan, 2000; Fischer et al., 2000; Morris et al., 2001). In these studies, fear conditioning has been indexed by measures such as changes in heart rate, skin conductance, or both, and imaged by using PET or fMRI to monitor changes in rCBF.
Guided by a large body of animal literature (Davis, 1992; Watkins et al., 1993; LeDoux, 1995; Li et al., 1996; LeDoux, 1998; Garcia et al., 1999; Guarraci et al., 1999; Holland and Gallagher, 1999; Hall et al., 2001; Neufeld and Mintz, 2001), one of the central foci of imaging studies of human fear conditioning has been the amygdala. Interestingly, although the first two PET studies to examine fear conditioning by using aversive stimulation identified areas such as thalamus, hypothalamus, central gray, anterior, and posterior cingulate cortex (Fredrikson et al., 1995) or orbitofrontal, prefrontal, frontal, and temporal cortices (Hugdahl et al., 1995), both failed to find the expected amygdala activation (for review see Buchel and Dolan, 2000). In these studies, rCBF data were compared before (habituation) and after (extinction) the acquisition (CS-US pairings) phase. In more recent PET studies (Morris et al., 1997, 1998), a differential conditioning paradigm has been used where facial expressions have been paired with loud noise. In the Morris et al. (1997) study, there was activation in the pulvinar and thalamus, which post hoc analyses revealed to be temporally correlated with the right amygdala. The post hoc analyses also revealed that the orbitofrontal cortex, hippocampus, and fusiform gyri also covaried positively with activation of the pulvinar. As in the eyelid conditioning studies reviewed above (e.g., Molchan et al., 1994; Schreurs et al., 1997b), this widespread correlation suggests a network of structures is involved in learning and memory and that the specific members of the network depend upon the type of task (McIntosh and Gonzalez-Lima, 1998).
In a follow-up study using differential conditioning of faces, Morris et al. (1998) showed that the amygdala was only activated during scans to the CS+ and not to scans of the CS-. Morris et al. (1998) also showed that the right amygdala was activated more than the left, if the CS+ was presented out of awareness. These data add to the debate about the role of awareness in human eyelid conditioning (Clark and Squire, 1998; Manns et al., 2000).
The most recent imaging studies of fear conditioning have used event-related fMRI to study the role of the amygdala (Buchel et al., 1998, 1999; Buchel and Dolan, 2000; Morris et al., 2001). These studies introduced and continue to use the partially reinforced differential conditioning paradigm described above in the Ramnani et al. (2000) study of eyelid conditioning. In the case of fear conditioning, a target tone or facial expression is paired with an aversive event such as a loud noise 50% of the time (CS+) and a second tone or facial expression is never followed by the aversive event (CS-). Skin conductance measures are usually obtained as a measure of conditioning. These experiments have been carried out during both a delay (Buchel et al., 1998) and trace (Buchel et al., 1999) conditioning paradigms. In delay conditioning the CS and US overlap whereas in trace conditioning there is a gap (i.e. the trace) between the end of the CS and the start of the US. In these studies, changes in anterior cingulate, insula, and amygdala were consistent with the previous fear conditioning studies. In addition, the trace study (Buchel et al., 1999) produced hemodynamic responses in the hippocampus, which is consistent with a large body of literature documenting the involvement of the hippocampus in trace conditioning (Port et al., 1985; Moyer et al., 1990; Kim et al., 1995; Woodruff-Pak and Papka, 1996; McEchron et al., 1998; Seager et al., 1999; Weiss et al., 1999). Finally, Morris et al. (2001) have been able to identify separate regions of the left amygdala that are involved during fear conditioning. A US-related lateral region, a CS+-related region in the ventral region, and a CS+-related region in the dorsal region that decreased as a function of continued differential conditioning.
The search for the biological basis of learning and memory has, until recently, been constrained by the limits of technology to classical anatomic and electrophysiologic studies. We have reviewed steady state animal experiments that image activity- and learning-dependent markers in fixed tissues, as well as dynamic humans studies based on PET and fMRI imaging of the intact brain that image changes in blood flow and metabolism during learning. The data prove that learning and memory can be imaged and that they involve a surprising conservation of mechanisms and the integrated networking of several structures and processes. In particular, the role of PKC in learning and memory has clearly been imaged in several animal model systems and shown to be a vital component in a cascade of memory-specific biochemical events. In both steady state animal imaging experiments and dynamic human functional imaging, the hippocampus and cerebellum are two structures that have consistently been shown to be part of a network that is involved in learning and memory. Nevertheless, it should be noted that the data obtained from steady state imaging in animal models are not isomorphic with those of functional imaging in human subjects. Indeed, although the animal literature seemed clear, early imaging studies of fear conditioning in humans failed to find changes in the amygdala. Moreover, data from human functional imaging studies are far from completely consistent between learning tasks or even within the same paradigm, as the data from eyelid conditioning experiments attest. Finally, without the appropriate control conditions, caution should also be exercised in an uncritical acceptance of any changes in functional activity representing changes involved in learning. Despite these limitations, it seems clear that the advent of imaging has provided an additional set of powerful tools that will help us to elucidate the neural substrates of learning and memory.
Dr. Alkon received his MD from Cornell University and is the Scientific Director of the Blanchette Rockefeller Neurosciences Institute and Professor of Neurology at West Virginia University. His research focuses on the molecular and biophysical bases of memory and memory dysfunction in neurological disorders, particularly Alzheimer disease. Dr. Schreurs received his PhD from the University of Iowa and is an Associate Professor in the Department of Physiology and in the Blanchette Rockefeller Neurosciences Institute. His research examines the behavioral laws and neural substrates of classical conditioning in human and animal models.