Social Networks, Cognition and Culture
Douglas R. White
Initial draft 8 for Blackwell Companion to Handbook of Cognitive Anthropology
Eds. D. Kronenfeld, G. Bennardo, V. De Munch, and M. Fischer
25 pp. (skip remainder) word count 9,170
Introduction.
Popper and Eccles (1977) debated the three-world problem that has confronted anthropology in terms of:
World 1: The physical world (and human brain and behavior in that world)
World 2: Mental activity and human consciousness.
World 3: Objective culture, "which is the creation of World 2 but takes on its own distinct and permanent existence."
The topics of this article and those of Read (2008) on formal empirical models confront the question of how these three worlds are related. How is it possible for Objective culture to take on a distinct and permanent existence? My Figure 1 brackets the three-world problem at two levels: that of the sciences (networks, cognition, culture) and how these play out at the individual level of brain, mind, and behavior or organism-environment linkage. Arguments between scientists such as neurophysiologist Damasio (2007) and philosophers like Gluck (2007), seemingly irreconcilable, fail to resolve these problems. For anthropology the interfaces between brain as a physical organ and the mind as a nonmaterial dynamical organized response pattern mediating the organism-environment (and self and other) interface creates problems of apparently incommensurate duality.
Yet, the sciences today are undergoing major transformations and re-synthesis. They in turn are affected by transformations in physics, biology and ecological psychology in dealing with complex systems and, in particular, the dynamics of complex systems. I want to address here how these new syntheses affect anthropology and those social sciences concerned with human cognition, culture, and networks. Cognitive anthropology (CA) is caught in a position of having to reconcile, somehow, individual cognition in the human brain with the existence of cultural patterns in terms of shared and meaningful symbols. I will not debate here the merits of CA as practiced in the 1960s and the solutions wrought by assuming that human thought operates through self-evident categories (“components” of meaning), and that symbols are somehow reducible to componential analysis, context, opposition, indexicality[1], and deixis.[2] Expressed in language, the convergence from raw sounds (phones) to phonemes emerges by memory and social convention (Skyrms 1996). Somehow, by analogy to phonemes, it was alleged that morphemes and word meanings in contrast sets emerge through components of meaning.
Because animals were recognized in laboratory experiments as making categorical judgments about colors, shapes, emotions, and the like, both material and “immaterial,” 1960s CA assumed that categories and their components must somehow impart meaning to the human mind. For humans, however, the assumption that mind operates through categories fails because humans also think relationally, as has been demonstrated experimentally (Hummel and Holyoak 2005, Penn, Holyoak, and Povinelli 2008). Thus, the first problem with CA in the 1960s, as shown in the first upper circle of Figure 1, was that meaning was seen as defined by categories, without consideration of relational cognition.
For Popper and Eccles, there is a fundamental problem in the relation of World 1 and World 2. How is mind (something immaterial) related to matter, i.e., to brain and the physics of behavior? Following the pioneers of cybernetics in the 1930s and the emergence of increasingly sophisticated electronic computers in the 40s and 50s, and given that the brain is chemo-electrical, and on the analogy of contrast sets in linguistics, the second problem of CA in the1960s was the assumption that the brain operated as a digital computer, a belief exaggerated by Lévi-Strauss in his 4-volume attribution of binary elementary logics to slow-moving “cold cultures” portrayed in Mythologiques as trying through repetition of mythic themes to resolve irresolvable contradictions. Oddly, this belief coincides with the Newtonian view of natural forces operating in four coordinates of space and time, as if time were ticking along and needing to be calculated digitally, in the brain, instead of analogically as flow. Einstein’s conception of light as constant flow that distorts gravity as velocity expands to light speed has impacted the social sciences only very slightly. Much contested, for example, but validated experimentally (Michaels and Carollo 1981) are Gibson’s (1950, 1955, 1966, 1972, 1979) findings on direct perception.
Thought (individual) Relations (social)
Cognition Networks
1960s: Categorical Perceptual flows
But also: of relations
Relational Behavior as epi-
Mind sodic relations
Brain Culture
Perception Emergent from
Emotion social interaction
Gibsonian in structurally
Behavior cohesive sets
–Environment
Fig. 1: The three-world problem at two levels: Individual and Social, Thought and Relations (Arrows represent a connective and circular logic, but influences are omnidirectional). Cognition and Culture are purposely not placed on the same plane to avoid the suggestion of reification of culture on a superorganic level of “group thought.” Culture, rather, is an emergent and not always singular or consistent distillation of practices in a cohesively organized group to which cognition may be directed, so there is a reflexive level of cognition in thinking about culture.
Summarizing experimental work in the Gibsonian tradition, Michaels and Carollo 1981:11-13) note that perception is "an ongoing activity of knowing the environment" that is not sliced problematically into past/present/future or into Newtonian time slices but a continuity of related parts of events, where "nowness" may or may not appear as a relevant phenomenal experience and is not axiomatic to a theory of knowing. It is relevant information that specifies the event and its parts, each of which last over time. A "whole event is perceived ... by a continuity of those 'parts'" and bounded by change to and from other perceived events, events in space-time and not metric coordinates. Information is not partitioned, it is coextensive with the event.
The experimentally validated solution to the mind-body conundrum, then, is that the embedded organism and its environment constitute a single system (Gibson 1986, West and King 1987, Swensen and Turvey 1991, Oyama 2000, Wagman and Miller 2003, Turvey 1991, 2009, Turvey and Shaw 1979, 1996). Events are bounded in perception by change in action that have networks of connected parts within events and recurrence across events. These views accord with Hutchins’ (1991, 1995) on the use of material anchors in studies of human cognition, i.e., the human-cognitive environment connection.
Time-series of events thus lend themselves (not emically but perceptually) to network coding and analysis. Such studies may be done at many different time scales. In our studies of kinship networks (see White and Johansen 2005 for an ethnographic example) there are intergenerational events such as marriage, childbirth, death, migration, and proximal interactions within these and other event boundaries and more micro or macro time-scales of event sequences. The network links among events and actors exhibit structural and dynamical patterns, including recurrences for which tools exist for studying complex dynamics (see Carollo and Moreno 2005 for methods), fractalities (White and Johansen 2005:136-137), and structural cohesion as a predictive network variable in relation to cultural formation.
Cohesive groups in networks. Cohesive blocks are found operationally in a manner that fits the basic conceptual form for the idea of the cohesion of groups, the way cohesion is perceived for groups, and the way that cohesion ties a group together both internally and by resistance to being dismembered. It also shows the way that networks provide a particular set of the degrees of freedom in how cohesive groups may relate to one another through overlap (e.g., membership in multiple communities) and through core-periphery subgroup hierarchies for levels of cohesion. This opens the way to the following hypothesis:
The Cohesion and Consensus Hypothesis: Levels and variations of cohesion within social networks for society as a whole and within its varying segments, which can be measured for level-specific cohesive blocks within networks, will tend to predict levels and variations in cultural consensus, provided that the connections that define the network have some positive perceptual relation to the subject(s) of cultural consensus.
A simple example was provided in the formulation of the measure of network cohesion by White and Harary (2001): they predicted how a Karate Club studied for two years by Zachary (1977) divided its membership between the club owner and the instructor, and the order of secession of those that followed the teacher to a new club. This has a cognitive dimension because (1) individuals must assess their relation to the owner and instructor in relation to others and decide with whom to affiliate or disconnect, and (2) they also assess who are their closer friends or allies in the network and who are more distant and how the two alternative leaders stand in relation to their closer friends. Personal attributes of the two leaders could not have made a better prediction. Defectors moving to the teachers side by breaking with those on the owner’s side do not follow an individual decision rule but follow a rule for the group: the tie that is least cohesive with the owner’s side is broken first, and in case of a tie for cohesion, the tie broken is more distant from the owner.
Moody and White (2003) showed (1) strong effects of levels of cohesion of individual students in their friendship blocks on their reports of attachment to high school and (2) how the cohesive strengths of co-memberships in the cohesive blocks of business alliances align with similarities in the choices of firms in their political party alliances of firms in party politics. In both cases, none of the other network or attribute variables outperformed the predictiveness of the cohesion measure.
Powell, White, Koput and Owen-Smith (2005), using the Moody-White measurement of structural cohesion, analyzed time-lagged effects from year to year of multiple variables in the choice of partners for strategic collaborations in the biotech industry and found they were strongly predicted by level of cohesion in the cohesive blocks to which potential partners belonged the year before. None of the other network or attribute variables outperformed the cohesion and diversity measures.
Multiconnectivity that implies structurally cohesive circuitry, especially for networks of organizations, is like a series of stacking blocks as shown in Fig. 2 (the children’s stacking-blocks game: with each successively smaller block stacked on a peg representing successively more k-cohesive groups), the top block representing the most cohesive subgraph of nodes in a network. The oddity is that blocks on different stacks may be part of a shared platform for their upper blocks (representing overlap for their lower blocks). The later Fig. 3 will show how individual blocs are defined by internal level of network-tie cohesiveness.
(Google image search “stacking blocks” p. 2)
Fig. 2: Stacking blocks, analogous to 3 cohesive hierarchies with overlaps of nodes in common
This kind of structure is difficult to envision because each k-component contains all blocks above a certain level and overlaps apply downwards to the blocks below. It is best stated abstractly as a mathematical definition for precisely bounded maximal subgraphs of a larger graph whose subgroups for levels k are found by algorithm, which give results easily perceived by humans when viewed in a suitable format. That is, such blocks are detectable in a sparse graph by keen visual perception after the nodes are pulled together if they are connected and pushed apart according to the length of the singular chains that connect them but which are not embedded in cohesive blocks. This is known as the spring embedding or FDP (Force-Directed Placement) visualization algorithm. A single longer chain both “connects” and “pushes apart” two nodes relative to nodes that are cohesively connected.
Differences between “integrated” single cohesive blocks and multiple “segregated” or “overlapping” cohesive blocks are illustrated in Fig. 3 (a)/(b) with two slightly different model networks: in one the ties are fully randomized (and thus are likely to be integrated), while in the other some strategic ties are added to create segregated (but in this case overlapping) organization clusters. Fig. 2 shows two graphs with 20 nodes, one with 38 random edges and another with 33 random edges plus 3 that force more cohesive complexity onto the graph. Fully random edges typically create embedded levels of “socially integrated” k-cohesion like a nest of Russian dolls. Fig. 3 (a) shows a sparse random graph with 20 nodes and 40 links, each link adding one degree to each of the two nodes linked, so the average degree per node is 4 edges (some have more than four and some less). Those that have degree four or more do are circled but no set of the 14 nodes with degree 4 forms a 4-component. There are 17 nodes, however, that form a 3-component. When edges are completely random, as in 3 (a), cohesive components form a single hierarchy, as in the case of the biotech network discussed above, where the graph is also cohesively integrated. In Fig. 3 (a) the 3-component is a subgraph of the 2-component, which nests in the largest (1-) component of all nodes that are connected.
The colors of nodes in Fig. 3 illustrate k-cores, which are uniquely defined in any network for the integers k=1,2,3,… where some of the higher k-cores may be empty. A k-core (for k=1,2,3,…) is the largest subgraph in which each node has degree k or more. Every k-component is a k-core but not every k-core is a k-component. In graph 3 (b) the two are equivalent for k=3,2,1, which is usual for perfectly random graphs and for “integrated” graphs with no tendency to form separate or overlapping k-components. The k-cores of a graph are easily computed, e.g., by Pajek, by iteratively deleting all nodes with less than degree k and recomputing degree, retaining those with k or more links. Like the measure of subgraph density, the use of k-cores (defined by Foster and Seidman, 1989, Seidman and Foster 1978) is often taken in network analysis as a measure of group cohesion, even though this usage is patently invalid (White and Harary 2001). For a more sophisticated use of k-cores as fingerprints of network structure, recognizing that cores may be disconnected, see Alvarez-Hamelin et al. 2006). A subgraph of two cliques (completely connected) may have 50% density and yet be disconnected, so that density is no measure of subgraph cohesion. Similarly, a k-core for any value of k with more than 2k nodes may be completely disconnected.