Most early MEG studies of basic vision [see review by Aine and Stephen (2002)] focused on examining properties (e.g., spatial or temporal frequency tuning properties) of a single visual area (e.g., primary visual cortex or V1) or have examined the retinotopic organization of visual areas (i.e., the point-to-point projection of visual field onto cortical areas). By carefully selecting stimulus parameters (e.g., color, size, motion), it is possible to identify and characterize different visual areas in the human brain similar to the methods employed in monkey studies. More recently, visual MEG studies have examined cognitive processes such as the representation of language in the brain, as well as memory and imagery. These studies rely less on the invasive results in monkeys and more on results from other functional neuroimaging methods such as PET and functional magnetic resonance imaging (fMRI) for corroboration [See review of fMRI visual studies in Courtney and Ungerleider (1997)].

Contemporary views of higher cognitive functions (e.g., cognitive neuroscience) date back to the late 1980s which emphasize that a number of different neural systems participate in the representation of an object or event (Kosslyn 1988; Squire 1986). A paradigm shift occurred away from the predominant view at that time that feature integration relies on convergent hierarchical processing, i.e., the visual system can be viewed as a series of processing stages that represent a progressive increase in complexity of neuronal representations that are dependent upon the output of preceding stages (De Yoe et al. 1994; Van Essen 1985; Van Essen and Maunsell 1983; Zeki 1978). While no investigator would argue that serial hierarchical processing does not occur within the visual system, a new emphasis was placed on “networks” or “systems.” The earliest example of a systems approach was the overwhelming evidence for the existence of at least two functionally specialized processing streams in the visual system (e.g., “dorsal” and “ventral”—see Fig. 1) operating in parallel (De Yoe and Van Essen 1988; Merigan and Maunsell 1993; Ungerleider 1995; Ungerleider and Mishkin 1982; Van Essen and Maunsell 1983). Basically, attributes of stimuli are not believed to be stored as a unified percept in a single cortical location; but rather, appear to be stored in a distributed cortical system in which information about specific features are stored close to the regions of cortex that mediate the perception of those features (Goldman-Rakic 1988; Mesulam 1998; Ungerleider 1995). Memory retrieval, therefore, means that a cue triggers a pattern of neural activity which is the same as the one elicited during the initial processing of the retrieved material (Alvarez and Squire 1994; Damasio 1989; Fuster 2001; Mesulam 1998; Mishkin 1982; Squire and Zola-Morgan 1991; Tulving 1995; Wheeler et al. 2000). Memories essentially consist of networks ranging from simple sensory memories or cell assemblies in sensory or parasensory areas to perceptual-motor associations consisting of reciprocal long-fiber connections linking perceptual memory networks of the posterior cortex with the prefrontal motor networks (Fuster 1997). The current results demonstrating consistent networks of activity during rest (e.g., default mode network or DMN) underline this view that brain activity during rest and task related activities involve broad cortical network activation.

Prefrontal cortex (PFC) was shown to be a key player in maintaining perceptual representations during working memory tasks and providing feedback to posterior cortex, thereby biasing activity in favor of behaviorally relevant stimuli (Baylis and Rolls 1987; Fuster 1973; Goldman-Rakic 1995). In monkeys, an elevated discharge in PFC during a delay interval was the most characteristic effect of the ‘sample’ stimulus (during encoding) on prefrontal cells suggesting that short-term memory basically consists of the continued facilitation of neural activity in cerebral structures recently engaged in sensory processing (Fuster 1973; Fuster and Jervey 1981; Miller et al. 1991; Richmond et al. 1983; Wilson et al. 1993). The monkey studies also showed that cooling of either PFC or inferotemporal cortex (ITC) interrupted loops of reverberating activity between them, a likely mechanism of the continued facilitation (Fuster 2001). Chafee and Goldman-Rakic (2000) found similar patterns of neuronal activity in PFC and parietal neurons in monkeys and demonstrated their interdependence via cortical cooling. These studies and others (Tomita et al. 1999) indicate that PFC-parietal and PFC-ITC regions share reciprocal projections and that these circuits are necessary for the transmission of receptive field properties and other dimensions of task-related activity when these areas are recruited to a common task. Our earliest memory study (Aine et al. 2003), designed to parallel the Delayed-Match-to-Sample (DMS) working memory studies conducted in monkeys, revealed strikingly similar results as those obtained in monkeys. Figure 2 (top row) shows similar time-courses for V1 in the monkey and humans. Walsh Stimuli (upper right) were used in both cases. The bottom row (left) shows an example from monkey area V4 when the monkey was attending a preferred stimulus versus attending a non-preferred stimulus. At the right, an effect of attention or working memory in humans also reveals elevated and sustained activity in several brain regions including the prefrontal region (Aine et al. 2003).

Cortical Feedback Connections. Although little is known about the physical nature or development of feedback connections, many believe that feedback projections from prefrontal cortex play a major role in modifying responses in lower-order brain regions during controlled or effortful processing, such as that required for attention and working memory tasks. Studies in nonhuman primates have revealed the existence of massive feedback projections that carry information from higher-order to lower-order regions. Although feedback connections can theoretically modulate early portions of the initial visual response (Hupe et al. 2001), most studies show that attentional modulations lag the earliest response (~250–300 ms in monkeys) (Haenny and Schiller 1988; Lamme et al. 1998; Mehta et al. 2000a; Motter 1994; Roelfsema et al. 1998; Seidemann and Newsome 1999). The late sustained activity allows information from feedback connections to be incorporated into the response to increase and sharpen the neural responses (Gilbert et al. 2001; Lamme and Roelfsema 2000). Our early attention studies using MEG were the first to suggest that attention can modulate V1 of humans via feedback from higher-order areas (Aine et al. 1995). These conclusions were based on the observations that: (1) attention-related effects in area V1 occurred later in time (~150 ms) than the earliest V1 activity and (2) this later attention-related activity showed a ~180º difference in net direction of current flow relative to the initial feedforward response around 80 ms. Previously, when ERPs were utilized to examine visual selective attention, concepts of “early in time” were often confused with “early levels of the visual system” suggesting that if attention had its effect early in time, then it must have occurred at the level of V1 or V2 and that if the attention-related effect occurred later in time, then it must have occurred at a higher level within the visual system (e.g., extrastriate regions). More recently, the attention effect in V1 was examined using both ERPs and fMRI (Di Russo et al. 2003; Martinez et al. 1999; Noesselt et al. 2002) and these studies supported the conclusions reached in Aine et al. (1995), as well as providing more definitive evidence. Additionally, fMRI studies (Brefczynski and DeYoe 1999; Tootell et al. 1998a) routinely show that attention can affect activity at the level of primary visual cortex and invasive studies show that it is via feedback influence which occurs later in time (Lamme and Roelfsema 2000; Mehta et al. 2000a, b). In general, MEG provides a unique tool for characterizing the spatio-temporal dynamics of neuronal activity that can even be used in certain circumstances (e.g., imagery) to differentiate feedforward activation sequences from feedback activity (Aine et al. 2003). Feedback connections are crucial for the establishment of neural circuits.

Feature Integration. How features and attributes of stimuli become integrated across widespread cortical regions has been an issue of intense interest and debate. Evidence indicates that (1) local field potentials (LFPs), which provide a measure of mainly postsynaptic dendritic responses, show strong sub-threshold synchrony of ongoing fluctuations in the cell’s membrane potentials (Lampl et al. 1999) and (2) coordinated sub-threshold excitability changes have been demonstrated to modulate local networks (Engel et al. 2001). From this perspective, induced oscillations are part of different cell assemblies that are activated to temporally bind different stimulus characteristics or to bind the activation of a system of distributed areas necessary for the task at hand. While the specific roles these rhythmic activities play are still debated (i.e., is it an epiphenomenon?), the existence of oscillatory activity is not (Salinas and Sejnowski 2001; Tallon-Baudry et al. 2004). But, as Tallon-Baudry et al. (2004) suggest, it is important to establish the behavioral relevance of oscillatory activity by showing that it is associated with behavior such as correct performance. Along these lines, Jensen and Tesche (2002) presented a list of digits similar to the original Sternberg design and found that MEG theta band activity over frontal regions increased parametrically with the number of items retained in working memory and there was stronger theta during the memory task compared to a control task. There was also a systematic increase in RT with increase in memory load. EEG and fMRI studies also suggest that an increase in frontal theta, associated with an increase in memory load, corresponds to a decrease in BOLD in DMN regions (Scheeringa et al. 2009). These studies suggest a functional role for oscillatory activity, but the conditions under which gamma, theta, beta or alpha activity is involved as well as their specific roles is still unclear. For example, does synchronization in theta band reflect episodic memory encoding (Klimesch 1999) or does it play a role in holding a stimulus in mind over the course of a brief delay (Lee et al. 2005)? MEG methods are uniquely suited for this exciting area of study.

As mentioned above, memories are formed through associations; brain regions or neural systems that are repeatedly active at the same time will tend to become associated (i.e., the principle of synchronous convergence) (Aertsen et al. 1989; Fuster 1997). Some investigators refer to the correlated activity across cortical areas as functional connectivity (Friston 1994) and others refer to it as temporal binding (Engel et al. 1992; Gray 1999; Roelfsema et al. 1997; Singer and Gray 1995). The former emphasizes the observed temporal correlation between two neurophysiological measurements from different parts of the brain (Gerstein and Perkel 1969) and the latter denotes the linking together of different features and attributes of stimuli through the selective synchronization of distributed neuronal activities (Bressler 1995; Gray 1999; Milner 1974; Singer and Gray 1995). Attention has been shown to enhance synchronization across different areas of cat brain (visual, parietal and motor cortex); with close to zero time lag (Roelfsema et al. 1997). While most studies in humans examine connectivity via coherence analysis using sensor measurements (sensor space), we have used our localized time-courses (source space) for cross-correlational analyses. Figure 3 shows a cross-correlation analysis (50–480 ms) conducted on the DMS data for three brain regions (medial occipital, medial temporal lobe or MTL, and PFC) and 4 participants. Similar to the Roelfsema study, attention or task relevance increases synchronicity across distributed cortical regions in healthy human participants (Aine et al. 2003). Activity in MTL and PFC regions covaried with each other (e.g., compare time-courses for one participant at left for MTL and PFC and cross-correlation plots for 4 participants at the right—green tracings). The maximum peak of the PFC and MTL correlation is near the zero lag while the maximum peak of the PFC and occipital correlation (shown in blue in the plots at the right) is not near zero. Lags and cross-correlation coefficients can be compared with behavioral performance measures to determine the relatedness of these brain regions. For a more thorough review of the functional organization of the visual system and of visual MEG studies in general, please refer to Aine and Stephen (2002).