Anthony Stigliani

Repetition of a stimulus typically leads to a reduction in neural response. This adaptation effect, sometimes known as repetition suppression or neural priming, can be observed both in individual neurons as well as fMRI voxels containing hundreds of thousands of neurons. When measured with fMRI, this repetition-related reduction in neural activity is known as fMRI-Adaptation (fMRI-A) and can be used to make inferences about the nature of neural representations and their sensitivity to various stimulus transformations (e.g., Grill-Spector et al. 1999). In fMRI-A designs, identical stimuli are presented repeatedly to participants, which is often accompanied by attenuation of fMRI signal relative to presentation of different stimuli. Then fMRI response to repetition of stimuli varied along one dimension is examined. If the neural representation in a given region is sensitive to changes in that dimension, the signal will recover to non-repeated levels, but if the representation is tolerant to changes in that dimension, the signal will remain adapted.

fMRI-A is a flexible method that has been applied to wide variety of topics, but some experimental designs are better suited for studying particular questions than others. The purpose of this wiki page is to describe the necessary components of fMRI-A designs as well as the advantages of different types of experimental designs.

Background

Figure 1. Response of an example IT neuron to first the presentation of stimulus A (left), the first presentation of stimulus B (center), and repeated presentation of stimulus A (right) (figure courtesy of Grill-Spector et al. 2006).

Repetition-related reductions in neural firing were first observed in electrophysiological recordings of neurons in macaque inferior temporal (IT) cortex (e.g., Li et al., 1993). Instead of illustrating a general reduction in neural firing when any two stimuli are shown in quick succession, response attenuation in this region was only observed only when identical visual stimuli were presented sequentially (see Figure 1). Similar effects of stimulus repetition are observed in the human brain with fMRI-A, and some regions in ventral temporal cortex show adapted responses even when the repeated stimulus undergoes various transformations such as changes in position in the visual field or retinal size (e.g., Grill-Spector et al., 1999). This raises the question of what counts as the "same" stimulus, but to some extent this depends on the particular question being asked. In studies of object recognition, individual items are typically object exemplars or faces, but fMRI-A has also observed in left inferior frontal cortex when different objects from the same category are shown sequentially as compared to when objects from different categories are presented in succession (Vuilleumier et al. 2002), which allows different hypotheses to be tested. Outside of this body of research, fMRI-A has been used to address a range of topics from the orientation tuning of neurons in visual cortex (Fang et al., 2005) to the representation of numerosity in parietal cortex (Piazza et al., 2004). The definition of an item varies greatly across these fMRI-A studies, but the power of the approach lies in its ability examine how different transformations affect the neural representation of individual items or stimuli as defined by the experimenter.

As illustrated in Figure 2, several neural mechanisms have been proposed to explain fMRI-A including general neural fatigue, sharpening of neural representations, and response facilitation resulting in reduced processing time (Grill-Spector et al., 2006), but recent evidence suggests that different mechanisms may operate across different brain regions and time scales of repetition (Weiner et al., 2010). In the fatigue model, all neurons that are responsive to a given stimulus show a reduction in firing rate when the stimulus is repeated. The sharpening model posits that fMRI-A reflects a sparser pattern of activation with repeated exposure to a stimulus, and the facilitation model holds that repetition leads to faster neural processing, which results in reduced fMRI response as a result of a shorter duration of signal accumulation. While these models implicitly assume that fMRI-A reflects reductions in the spiking rates of individual neurons, both the inputs and outputs to a brain region contribute to BOLD signal, and it is important to consider that some repetition-related response reductions could be inherited from earlier regions (see Logothetis, 2008).

Figure 2. Neural models of fMRI-A. Different neural mechanisms have been proposed to explain repetition-related reductions in fMRI signal. The fatigue model posits a general reduction in neural response with stimulus repetition. The sharpening model holds that repetition results in activation of a sparser neural representation, and the facilitation model attributes fMRI-A to a reduction in processing time (figure adapted from Grill-Spector et al., 2006).

Measuring fMRI-A

Three measurements must be made to make inferences about the nature of neural representations using fMRI-A:

1. Identical condition: Response when an identical stimulus is repeated
2. Different condition: Response when different stimuli are presented
3. Transformation condition Response when the same stimulus is repeated but varied along one dimension

In a typical fMRI-A design, a baseline fMRI-A effect is defined as reduced fMRI signal in the identical condition compared to the different condition. Response to the transformation condition is then used to assess the sensitivity of a neural representation to the varied dimension.

Measuring a Dynamic Range of fMRI-A

BOLD response to the first two conditions must be measured in order to define a dynamic range in which to compare response to the transformation condition. If fMRI signal in a voxel is reduced when an identical stimulus is repeated relative to presentation of different stimuli, this suggests that underlying neural representation in this region is stimulus-specific. This inference is based on the notion that presentation of different stimuli leads to activation of different populations of neurons in a voxel, and therefore the neither population becomes adapted. However, if there are populations of neurons that respond selectively to the repeated stimulus, subsequent presentations of the same item will activate the same population of neurons repeatedly resulting in an adapted fMRI response.

Assessing Sensitivity and Invariance

After measuring the baseline fMRI-A effect in a given voxel, fMRI response to the transformation condition is compared to the identical and different conditions. If the neural representation of a stimulus is sensitive to the varied dimension, fMRI signal should be comparable to conditions in which different stimuli are presented because different populations of neurons are activated when this feature of the stimulus is manipulated. If the representation is invariant or tolerant to changes in the dimension of interest, fMRI signal should remain at levels associated with repetition of an identical stimulus because the neurons are figuratively blind to changes in the varied dimension and the same population responds to both variants of the stimulus. While comparisons between these three (or more) conditions are proposed to assess neural population tuning at sub-voxel resolution, it is important to realize that voxels likely contain a mixture of neurons that are sensitive and tolerant to changes in the dimension of interest. Nevertheless, adapted (or recovered) responses can be interpreted to reflect the response profile of the majority of stimulus-responsive neurons in a voxel.

An Example from Face Perception

As an example of how fMRI-A can be used to assess the nature of neural representations, imagine two hypothetical populations of neurons involved in representing a specific face. In one population, different sets of neurons are activated when the face is viewed from different angles (viewpoint-specific neurons). In the other population, the same set of neurons responds across all viewpoints (viewpoint-invariant neurons).

Figure adapted from Grill-Spector & Malach (2001)

When the same image of the face is shown repeatedly, reduced fMRI response is observed in both scenarios because the same population of neurons is activated across the repeated presentations of the same image (compared to fMRI signal in an epoch of different faces).

Figure adapted from Grill-Spector & Malach (2001)

When the same face is shown from different views, fMRI response would differ depending upon the sensitivity of the underlying neural representation to changes in viewpoint. If the majority of neurons in a voxel respond to the face in a viewpoint-specific manner, the fMRI signal in that voxel would show a recovery from adaptation because different populations of neurons are activated by the different views. On the other hand, if most neurons responded in a viewpoint-invariant manner, fMRI signal should remain adapted at levels observed in epochs of identical images.

Figure adapted from Grill-Spector & Malach (2001)

Experimental Design

fMRI-A has been measured using a variety of stimulus presentation parameters and experimental designs, each of which has advantages and disadvantages. Some studies use blocked designs in which adaptation is observed across several presentations of the same stimulus, while others use event-related designs in which adaptation is observed across two presentations of the same stimulus. In general, block designs produce larger effect sizes (Grill-Spector et al., 2006), but event-related designs are more appropriate in certain circumstances such as when the predictability or full counterbalancing of stimuli is important.

Block Design

In blocked fMRI-A designs, a series of stimuli belonging to the the same experimental condition are presented in groupings that are ideally counterbalanced across conditions over the course of the experiment. In the aforementioned experiment examining the the sensitivity of neural representations to changes in viewpoint, for example, an identical condition block would consist of several sequential presentations of the same face shown from the same viewpoint each time. Transformation conditions blocks would consist of several presentations of the same face shown from different viewpoints, and different condition blocks would contain several images of different faces.

Advantages

• Generally produces larger effect sizes (see Grill-Spector et al., 2006)
• Relatively simple to generate stimulus sequences and model fMRI-A effects

Disadvantages

• Stimuli are somewhat predictable based upon the first few stimulus presentations in a block
• More difficult to implement parametric designs in many cases

Event-Related Designs

Paired adaptation design

In paired adaptation designs, fMRI response to pairs of stimuli are the units of analysis which allows both stimuli to be presented in one TR and modeled as a single experimental event. With regard to the aforementioned face perception study, identical condition events would consist of two sequential presentations of the same image. Transformation condition events would consist of sequential presentations of the same face from different viewpoints, and different condition events would consist of images of two different faces.

Advantages

• Stimuli are less predictable than in block designs
• More conditions of interest can often be included in the experiment because individual epochs have shorter durations
• More suitable for implementing parametric designs

Disadvantages

• Smaller effect sizes

Lagged adaptation design

In lagged adaptation designs, pairs of stimuli are repeated over the course of a run (or in different runs) but not on sequential trials. There is some evidence that different neural mechanisms account for fMRI-A observed with many intervening stimuli (Weiner et al., 2010). In this type of design, trials consisting of a single stimulus presentation are modeled as novel or repeated, and the number of intervening stimuli must be balanced across different stimulus conditions.

Advantages

• Stimuli are completely unpredictable
• Allows potentially different sources of fMRI-A to be measured by varying the time between stimulus repetitions

Disadvantages

• Smaller effect sizes
• Stimulus sequences are difficult to generate because the number of intervening stimuli must be balanced across experimental conditions

Continuous-carryover design

Continuous carry-over designs use serially balanced stimulus sequences which simultaneously counterbalance main effects and first-order carry-over effects for all stimuli included in the sequence. This allows fMRI-A and multi-voxel pattern analysis (MVPA) to be performed in the same dataset because the first order carry-over effects measured with fMRI-A are independent of the counterbalanced main effects required to perform MVPA. This type of design is well suited to model fMRI-A when the stimuli of interest can be modeled to fill a continuous dimensional space such as shape or color. For example, a continuous carry-over design was recently used to distinguish between brain regions that represent color as a continuous dimension and regions that represent colors in a more categorical manner (Persichetti et al., 2013).

Advantages

• Allows fMRI-A and MVPA to be performed in the same dataset
• Suitable for implementing parametric designs

Disadvantages

• Smaller effect sizes
• Serial balancing of all stimuli makes design less efficient for addressing specific questions using fMRI-A
• A relatively small number of stimuli may be used because all images are counterbalanced for first order carry-over effects

References

1. Fang, F., Murray, S.O., Kersten, D., & He, S. (2005). Orientation-tuned fMRI adaptation in human visual cortex. Journal of Neurophysiology, 94(6), 4188-4195.
   
2. Grill-Spector, K., Kushnir, T., Edelman, S., Avidan, G., Itzchak, Y., & Malach, R. (1999). Differential processing of objects under various viewing conditions in the human lateral occipital complex. Neuron, 24(1), 187-203.
   
3. Grill-Spector, K., & Malach, R. (2001). fMR-adaptation: a tool for studying the functional properties of human cortical neurons. Acta Psychologica, 107(1-3), 293-321.
   
4. Grill-Spector, K., Henson, R., & Martin, A. (2006). Repetition and the brain: neural models of stimulus-specific effects. Trends in Cognitive Science, 10(1), 14-23.
   
5. Li, L., Miller, E.K., & Desimone, R. (1993). The representation of stimulus familiarity in anterior inferior temporal cortex. Journal of Neurophysiology, 69(9), 1918-1929.
   
6. Logothetis, N.K. (2008). What we can do what we cannot do with fMRI. Nature, 453(7179), 869-878.
   
7. Piazza, M., Izard, V., Pinel, P., Le Bihan, D., & Dehaene, S. (2004). Tuning curves for approximate numerosity in the human intraparietal sulcus. Neuron, 44(3), 547-555.
   
8. Vuilleumier, P., Henson, R.N., Driver, J., & Dolan, R.J. (2002). Multiple levels of visual object constancy revealed by event-related fMRI of repetition priming. Nature Neuroscience, 5(2), 491-499.
   
9. Weiner, K.S., Sayres, R., Vinberg, J., & Grill-Spector, K. (2010). fMRI-adaptation and category selectivity in human ventral temporal cortex: regional differences across time scales. Journal of Neurophysiology, 103(6), 3349-3365.