Distributed Hierarchical Processing in the Primate Cerebral Cortex
Summary (8 min read)
A Cortical Map
- The authors primary format for illustrating the location of different visual areas involves the use of 2-D cortical maps that are generated from contours of layer 4 in a series of regularly spaced histological sections (Van Essen and Zeki, 1978; Van Essen and Maunsell, 1980) .
- That map was not especially accurate, however, because of the large and somewhat nonuniform spacing between sections, and no scale was provided.
- In addition to the obvious cut that surrounds area VI (the elliptical region on the left), there are 2 smaller discontinuities, one along the ventrolateral side of the frontal lobe (upper right), and the other at the temporal pole (lower right).
- The remainder of the perimeter of the map represents intrinsic borders between the cortex and various noncortical structures (e.g., the dentate gyrus, amygdalar nuclei, and corpus callosum).
Visual Areas
- Altogether, there are 32 separate neocortical areas that are implicated in visual processing, based on the occurrence of visually responsive neurons and/or the presence of major inputs from known visual areas.
- The authors have drawn a distinction between 25 areas that appear to be predominantly or exclusively visual and another 7 neocortical areas that are less intimately linked to vision and will be considered visual-association areas.
- The criteria used in identifying these areas are discussed in detail below.
- A key that links these abbreviations to the standard text citations is given in the table notes.
- In other cases, the alternative scheme is even more fine grained (e.g., POa-i and POa-e vs. LIP).
Surface Area
- Measurements of the surface area of different regions of the cortical map (Table 2 ) provide useful information about the absolute and relative amounts of cortical machinery devoted to different types of processing.
- Besides the neocortex, there are 245 mm 2 of the hippocampus proper (fields CA1 and CA3, the subiculum and the prosubiculum), 120 mm 2 of paleocortex (pyriform and periamygdaloid cortex), and 270 mm 2 of transitional cortex [entorhinal cortex (ER), periallocortex, parasubiculum, presubiculum, and prostriate cortex].
- There are also inaccuracies in their transposition of areal boundaries defined in other studies onto the particular hemisphere used for this map (see above).
- These are hard to quantify, but they probably reflect errors of 50% or more for some areas.
- Areas with sharply defined borders such as VI and MT show roughly 2-fold individual variability in surface area (Van Essen et al., 1981 ,1984) , and it seems likely that this range will be applicable to most, if not all, cortical areas.
Connectivity
- Nearly all of the areas included in this scheme can be distinguished on the basis of their overall pattern of connectivity, and for many, this is the primary basis for identification.
- The authors have included entries for both coarser and finer subdivisions when appropriate in the table.
- Many areas, particularly the recently defined ones, have yet to be studied in detail; hence, their description of the connectional pattern is surely far from complete.
- Each row shows whether the area listed on the left sends outputs to the areas listed along the top.
Specific Visual Areas
- In order to put the current map into perspective, it is useful to comment on the layout of specific visual areas, with emphasis on recently identified areas and areas for which uncertainties in identification persist.
- These are surrounded anteriorly by a collection of smaller areas, 3 of which have been mapped in some detail (MT, V3, and VP), and the remainder of which are less well characterized (V3A, V4t, and VOT).
- This region includes 3 areas (PO, PIP, and DP) situated posteriorly, 5 areas (7a, LIP, VIP, MIP, and MDP) situated more anteriorly and arranged in a lateral-to-medial swath that adjoins the somatosensory cortex, and 2 areas (MSTd and MST1) within the dorsal part of the superior temporal sulcus (STS).
- Overall, it remains unclear whether this heavily myelinated strip should be considered part of VIP or LIP or as a distinct area unto itself.
- Based on the heterogeneous pattern of connectivity with parietal and temporal areas, there are probably distinct subdivisions within area 46 (Goldman-Rakic, 1988; Barbas and Pandya, 1989; Cavada and Goldman-Rakic, 1989b; Seltzer and Pandya, 1989a) , but a coherent scheme for subdividing it has yet to emerge.
Reciprocity and Distributed Connectivity
- This tabulation also provides a useful framework for discussing other important principles concerning the numbers and patterns of connections among different areas.
- The degree to which this relationship holds is reflected in the symmetry of Table 3 about the diagonal axis (shaded boxes).
- A lower bound on the overall degree of connectivity can be set from the overall number of 305 identified pathways linking 32 visually related neocortical areas.
- If each area has an average of 27 connections, as found for well-studied areas, the connectivity level would exceed 40% of the theoretical limit (432/992).
- The fraction of pathways that are "robust," in the sense of showing heavy labeling when analyzed with conventional tracers, may be only 30%-50% of the total number of identified connections.
Hierarchical Relationships in the Visual Cortex
- The possibility that the visual cortex might operate by a strictly serial processing scheme can be ruled out just from knowing the multiplicity of connections per area and the near ubiquity of reciprocal connections.
- One hypothesis is that cortical areas are hierarchically organized in some very well-defined sense, with each area occupying a specific position in relationship to all other areas, but with more than 1 area allowed to occupy a given hierarchical level.
- Others are less well defined, and there may be basic uncertainties as to who ranks above whom in various interactions.
- It is worth noting in general terms that information flow in a hierarchical system (1) can go in both directions (upwards and downwards), (2) can skip over intermediate levels to go directly from a low to a high level, and (3) can travel in parallel through multiple, functionally distinct channels.
- The number of identified pathways for which useful laminar information is available has more than tripled in the past 5 years.
Criteria
- The authors revised criteria for identifying hierarchical relationships are illustrated schematically in Figure 3 .
- The different patterns are arranged to show the laminar distributions of cells of origin and axonal terminations that the authors consider to be indicative of ascending (A, upper row), lateral (L, central row), and descending (D, bottom row) pathways.
- In one pattern (F), terminations are densest in layer 4, though they may also be prominent in layer 3 and other layers, as well.
- Occasionally, patterns are encountered that involve primarily superficial layers (e.g., layers 1 and 2 in the projection from V4 to VI; predominantly layer 3 in the projection from AITd to FEF).
- Such B-F combinations invalidate one of the initial assumptions about feedforward pathways, but they do not necessarily invalidate the notion of hierarchical organization.
A Database for Anatyxtng Hierarchical Relationships
- The authors goal in this section is to apply the scheme illustrated in Figure 3 as objectively and rigorously as possible to the analysis of hierarchical relationships in the visual cortex, while taking into account the uncertainties and qualifications that are associated with some of the experimental data.
- The ease and reliability with which such data could be related to their partitioning scheme varied widely and depended to a large extent on how much detailed information was given about the pathways under consideration.
- As with the areal assignments, the determination of laminar patterns associated with each pathway was often difficult, depending on the nature and extent of published information available.
- For this reason, their computerized database included, in addition to the relevant publications, a listing of their specific page numbers and figure numbers that are particularly informative about laminar patterns.
- With respect to determining hierarchical relationships, however, the presence of a mixed result (S/B or I/B) in a single pathway does not represent an inherent conflict, because a bilaminar pattern is consistent with all possibilities, ascending, descending, or lateral.
Levels crossed
- This table shows connections among visual conical areas listed in Table 1 .
- The authors suspect that this bias for C/F and C/M patterns is not a coincidence and that it may be important for understanding the significance of mixed or intermediate termination patterns (see below).
- To avoid logical inconsistencies, each area must be placed above all areas from which it receives ascending connections and/or sends descending connections.
- Areas MIP and MDP have been placed at the fifth hierarchical level, even though the connections known for both areas are ambiguous (bilaminar retrograde labeling) and would technically be consistent with placement at any lower level.
- It is notable that all 3 of these inconsistencies involve relationships that were already questionable from an earlier stage of the analysis.
Significance of Hierarchical Irregularities
- Their presence raises the issue of whether the cortex is inherently only a' 'quasi-hierarchical" structure that contains a significant number (perhaps 10%) of bona fide irregularities and exceptions to any set of criteria that can be devised.
- The anatomical data on which their analysis is based are often fuzzy and replete with uncertainties of one or another type.
- Thus, it would have defied the odds if every single one of the 305 pathways had fit precisely into an orderly hierarchy.
- If one suspects that the underlying biology is extremely orderly, one would predia that the apparent discrepancies listed in Table 6 will largely disappear upon careful reexamination, thereby improving the overall fit to the hierarchy.
- The authors found that resolution limits made it impractical to flag these special cases by distinctive colors in the figure, but they can nonetheless be readily tracked down with reference to Tables 3 and 5 .
Number of Levels Traversed
- While some pathways link areas at the same or immediately adjacent hierarchical levels, the majority of pathways traverse more than 1 level.
- Figure 6 shows that there is an interesting difference along these lines.
- Nonetheless, it is apparent that the more specific unilaminar projections, on average, traverse more hierarchical levels than do the bilaminar projections: 2.68 levels for I patterns, 2.76 levels for S patterns, and 1.71 levels for B patterns, excluding all of the lateral pathways.
- A related set of questions arises when considering connectivity patterns and hierarchical relationships among adjacent visual areas, that is, ones that share a common boundary in the intact cortex.
- There are numerous examples of neighboring areas separated by 2 or 3 levels (e.g., V2/V4, VP/V4, and PIP/VIP).
Hierarchical Relationships in Other Regions and in Other Species
- The visual cortex has extensive connections with a variety of nonvisual areas, both cortical and subcortical.
- This allows us to link the visual hierarchy with a somatosensory hierarchy that will be discussed below.
- The analysis is less straightforward for entorhinal cortex, a complex of several small areas (Amaral et al., 1987) all having a transitional architecture that lacks the cell-dense layer 4 characteristic of most neocortical areas.
- The architecture and connectivity of the hippocampal complex is radically different from the neocortical areas discussed above (cf. Swanson et al., 1987) .
- Hence, it should not be surprising that a modified set of criteria would be necessary for making any hierarchical assignments.
Somatosensory and Motor Cortex
- The notion that forward and feedback connections can be used to delineate hierarchical relationships is nearly as old for the somatosensory cortex as it is for the visual cortex.
- As in the visual system, reciprocity of connections between areas appears to be a general rule, but there are several possible exceptions, including pathways from 7b to 1, SII to 4, and granular insular (Ig) to dysgranular insular (Id) that apparently lack connections in the reverse direction.
- By this point, it should not be surprising to find that there are a few irregularities that must be addressed.
- Figure 7 shows the somatosensory-motor hierarchy that results from the systematic application of the pairwise hierarchical assignments contained in Table 8 .
- In brief, this hierarchy starts with areas 3a and 3b at the bottom and extends in successive stages through areas 1, 2, 5, retroinsular (Ri), SII, 7b, Ig, and Id.
Auditory Cortex
- In the auditory system, Galaburda and Pandya (1983) analyzed connections among 12 cytoarchitectonic areas that they identified within the superior temporal gyrus and supratemporal plane of the lateral sulcus.
- These areas were grouped into 4 rostrocaudally aligned triplets of "root," "core," and "belt" areas.
- With anterograde labeling, they found that the feedbacktype pattern was generally strongest in layer 1, but otherwise conformed to the F, C, and M description that the authors have used.
- They reported that rostral-to-caudal projections tended to be of the descending pattern, and that caudal-to-rostral projections tended to be ascending in some cases but columnar in others.
- One small piece of evidence in further support of an auditory hierarchy comes from a single tracer injection in the postauditory area (Pa), which demonstrated descending projections to Al and ascending connections to a different auditory area (Friedman et al., 1986) .
Other Cortical Regions
- The remaining regions of the neocortex yet to be incorporated into their analysis include much of the frontal lobe (orbitofrontal, lateral prefrontal, dorsal prefrontal, and medial prefrontal), as well as cingulate, retrosplenial, and insular regions.
- There is not a great deal of information about the specific laminar patterns for pathways to and from precisely defined areas in these regions.
- One striking finding is that large paired injections centered in areas 7a and 46 led to interdigitating columnar patterns of terminations in some regions (e.g., cingulate cortex and orbitofrontal cortex), even though the same injections contributed to complementary (ascending and descending) patterns in other regions, such as the STS (Selemon and Goldman-Rakic, 1988) .
- This trend may instead simply represent the greater uncertainty and ambiguity about many of the high-level assignments.
- Each of these regions receives direct inputs from the olfactory bulb that, as already noted, terminate preferentially in superficial layers of cortex (Turner et al., 1978) .
Subcortical Projections
- All visual areas that have been appropriately examined have extensive connections with a variety of subcortical structures.
- Indeed, it would not be surprising if the sheer number of corticosubcortical pathways exceeds that of the corticocortical pathways analyzed in this article.
- The projections from cortex to different pulvinar subdivisions originate predominantly from layer 5, and the reciprocal projections from the pulvinar terminate most heavily in layers 4 and 3 of the extrastriate cortex (Lund et al., 1975; Benevento and Rezak, 1976; Ogren and Hendrickson, 1977) .
- Interestingly, however, the pulvinar projection to VI terminates mainly in superficial layers, even though the reciprocal pathway originates from layer 5, just as for extrastriate areas (Rezak and Benevento, 1979) .
- This is consistent with the amygdala being at a well-defined level just below TF and TH.
Other Species
- There is also a considerable body of information about laminar connectivity patterns in other species.
- The authors have already discussed the need in the primate cortex to treat bilaminar retrograde labeling patterns as completely ambiguous with regard to hierarchical assignments.
- By applying their revised criteria to connectivity patterns described in Symonds and Rosenquist (1984a,b) for visual cortex in the cat, the authors have constructed an orderly hierarchy that involves 62 connections among 16 areas organized into 8 levels (Fig. 8 ).
- They are followed by 2 levels, each containing numerous entries (areas PLLS, PMLS, SVA, and ALG at the fourth level and areas AMLS, ALLS, DLS, VLS, 21a, and 20b at the fifth level).
- Much remains to be done in order to resolve the modest number of apparent discrepancies and to ascertain just how generally this hypothesis applies across systems and species.
Intertwined Processing Streams in the Visual Cortex
- The notion of parallel processing streams in the visual system has received considerable attention during the past decade and is the topic of several recent reviews (e.g., Livingstone and Hubel, 1987b; Maunsell and Newsome, 1987; DeYoe and Van Essen, 1988; Lennie et al., 1990) .
- The central issue the authors wish to address in the remainder of this article is the relationship between the low-level M and P streams that originate in the retina and the high-level streams associated with areas in the temporal and parietal lobes (Ungerleider and Mishkin, 1982; Desimone and Ungerleider, 1989) .
- A second form of cross talk occurs in the ascending connections between areas.
- Likewise, MT projects heavily to the parietal cortex (directly to VIP and indirectly via MSTd and MST1), but it also has indirect connections with inferotemporal areas via FST and V4.
Single Neuron Connectivity
- Thus far, the authors have concentrated on the connections of entire areas or of layers and compartments within areas, without addressing the issue of heterogeneity among the individual neurons that make up a layer or an area.
- Most of what the authors know about this issue comes from a relatively small number of double-retrograde-labeling studies in cats and monkeys, in which tracers are injected into topographically corresponding portions of 2 different areas (cf. Kennedy and Bullier, 1985; Bullier and Kennedy, 1987) .
- In general, diis approach reveals a significant number of doubly labeled cells, signifying that individual neurons can indeed have collaterals projecting to more than 1 area.
- The average number of target areas per cortically projecting neuron could plausibly be well under or well over 2.
- In the cat, there is evidence that this number is greater for descending pathways than for ascending pathways, and that some cells can even contribute simultaneously to both directions, by making both an ascending and a descending connection (Bullier et al., 1984; Bullier and Kennedy, 1987) .
Functional Implications
- The authors have concentrated in this study primarily on an anatomical analysis that suggests 5 key principles of primate cortical organization: (1) a large number of visual areas, (2) highly distributed connectivity among areas, (3) reciprocity of connections, (4) hierarchical organization, and (5) distinct, yet intertwined, processing streams.
- The authors now comment on what these principles might signify for understanding the functions of different visual areas.
Distributed Hierarchical Processing
- The hierarchical scheme for visual cortex that the authors have presented is grounded explicitly on anatomical criteria.
- That situation is now changing, and a few of the more notable examples are worth explicit mention: (1) Many cells in V2, but not in VI, are responsive to patterns that elicit percepts of subjective contours in human observers (Peterhans and von der Heydt, 1989; von der Heydt and Peterhans, 1989) .
- These and other examples support the notion that higher stages of the cortical hierarchy represent more advanced levels of processing.
- The physiological properties discussed thus far (increases in classical receptive field size and more advanced receptive field selectivities) may largely reflect the contributions of ascending pathways and of circuitry intrinsic to each area.
- The physiological properties of any given cortical neuron will, in general, reflect many descending as well as ascending influences.
Functionality of Processing Streams
- The M stream contains a high incidence of cells selective for direction of motion and for binocular disparity, suggesting that it is heavily involved in the analysis of motion and depth.
- First, consider what sources of information are useful for signaling object motion.
- The effects of selectively lesioning the M and P layers of the LGN on specific behavioral tasks provide support for the notion that M and P channels each contribute to multiple aspects of perception (Schiller and Logothetis, 1990; Schiller et al. 1990; Merigan et al., 1991) .
- In a more general sense, there appears to be a complex, but orderly, relationship between low-level sensory cues (e.g., orientation, velocity, disparity, and spectral composition), high-level aspects of perception (e.g., perception of shape, surface qualities, and spatial relationships), and the processing streams that generate one from the other (DeYoe and Van Essen, 1988) .
Notes
- Two such enamel-painted, plaster-coated, styrofoam models were available, one at 3 times life size and the other at a scale of 9-fold.
- Boundaries of individual cortical areas identified in the studies indicated in the text and in Table 1 were marked onto the brain model, mainly on the basis of the relationship to various geographical landmarks.
- Once the physical model had been marked, the various areal boundaries were transposed to outlines of the sections on which the model was based.
- The next step was to transpose boundaries to sections of the brain from which the cortical map was made.
- The manually generated map, complete with areal boundaries, was optically scanned and used as a template for creating the color map with the CANVAS program on a Macintosh II computer.
Did you find this useful? Give us your feedback
Citations
[...]
46,982 citations
[...]
38,208 citations
14,635 citations
Additional excerpts
...…Rossi, & Desimone, 2005; Connor, Brincat, & Pasupathy, 2007; Desimone, Albright, Gross, & Bruce, 1984; DiCarlo, Zoccolan, & Rust, 2012; Felleman & Van Essen, 1991; Hubel & Wiesel, 1968; Hung, Kreiman, Poggio, & DiCarlo, 2005; Kobatake & Tanaka, 1994; Kriegeskorte et al., 2008; Lennie & Movshon,…...
[...]
9,700 citations
Cites methods from "Distributed Hierarchical Processing..."
...The pathways identified by these methodologies have formed the basis for the systematic collation of species-specific anatomical connection matrices, including those for the macaque visual corte...
[...]
9,599 citations
References
12,923 citations
6,388 citations
3,185 citations
"Distributed Hierarchical Processing..." refers background in this paper
...In the visual cortex, each processing stream maintains a distinct profile of receptive field characteristics (for reviews, see DeYoe and Van Essen, 1988; Livingstone and Hubel, 1988)....
[...]
2,612 citations
1,610 citations
"Distributed Hierarchical Processing..." refers background in this paper
...…contributions include inputs from other sensory modalities (especially auditory and somatosensory ), visuomotor activity (i.e., related to eye movements), and attentional or cognitive influences (cf. Andersen, 1987; Maunsell and Newsome, 1987; Goldman-Rakic, 1988; Desimone and Ungerleider, 1989)....
[...]