Hierarchical Perceptual Organization

Ruth Kimchi

University of Haifa

Leuven, Belgium 2014

Marco Polo describes a bridge, stone by

stone.

“But which is the stone that supports the

bridge?” Kublai Khan asks.

“The bridge is not supported by one stone or

another,” Marco answers, “but by the line of

the arch that they form.”

Kublai Kahn remains silent, reflecting. Then

he adds: “Why do you speak to me of the

stones? It is only the arch that matters to me.”

Polo answers: “Without stones there is no

arch.” Invisible Cities by Italo Calvino, 1972/1974, p. 82

Perceptual Relations between Parts and Wholes

• Visual objects can be viewed as hierarchical structure of parts and wholes.

• Structuralist: rooted firmly in British Empiricism, perceptions are constructed from atoms of elementary, unrelated local sensations that are unified by associations due to spatial and temporal contiguity.

• Gestalt: rejected both atomism and associationism. According to its doctrine of holism, a specific sensory whole is qualitatively different from the complex that one might predict by considering only its parts, and the quality of a part depends upon the whole in which it is embedded (Köhler, 1930/1971; Wertheimer, 1923/1938).

Some of the modern attempts to grapple with the issue part-whole relationships in human perception:

• Global precedence

• Primacy of holistic/configural properties

Global Precedence• The global precedence hypothesis (Navon, 1977):

perceptual processing proceeds from the global structure towards analysis of more local details.

• The framework: A visual object can be viewed as a hierarchy of

parts and subparts interrelated by spatial relationships

The globality of a visual property corresponds to the place it occupies in the hierarchy: Properties at the top of the hierarchy are more global than those at the bottom, which in turn are more local.

According to the global precedence hypothesis, global properties of a visual object are processed first, followed by analysis of more local properties.

Consistent Inconsistent350

400

450

500

550

600

650

700

Hierarchical Condition

RT

(m

se

c.)

Global

Local

Global-Local Paradigm

After Navon

(1977)

• Global advantage: faster identification of the global letter than the local letter

• Global-to-local interference: disruptive influence from irrelevant global conflicting information on local identification

Global advantage: boundary conditions

• Overall visual angle (e.g., Kinchla & Wolfe, 1979)

• Eccentricity of presentation (e.g., Pomerantz, 1983)

• Spatial certainty (e.g., Lamb & Robertson, 1988)

• Sparsity of elements (e.g., martin, 1979)

• Exposure duration (e.g., Paquet & Merikle, 1984)

• Goodness/meaningfulness of elements (e.g., Poirel et al., 2006)

• Number & relative size of elements (e.g., Kimchi, 1988, 1998)

The source of global advantage

• Perceptual (e.g., Navon, 1977, 1991)

• A sensory mechanism – faster processing of low spatial frequencies than high spatial frequencies (e. g., Shulman & Wilson, 1987)

• Post-perceptual (e.g., Miller, 1981; Ward, 1982)

Brain localization

• Right hemisphere biased toward global processing and the left hemisphere biased toward local processing (e.g., Robertson et al, 1993)

– relation between spatial frequency processing and global and local perception (e.g., Ivry & Robertson, 1998)

– Saliency of the stimulus: right hemisphere biased toward more salient objects and left hemisphere biased toward less salient objects (Mevorach, Humphreys, & Shalev, 2006a, 2006b)

– Integrating shape and level information: right hemisphere - binding shapes to the global level, left hemisphere - binding shapes to the local level (Hubner &

Volberg, 2005)

From a perceptual organization perspective:

• Perceived Hierarchical Structure (Kimchi & Palmer, 1982)

• Microgenesis of the organization of hierarchical structure (Kimchi, 1998)

Critical role of number and relative size of elements

Kimchi &Palmer, 1982

Kimchi &Palmer, 1982

Microgenetic Analysis of Perceptual Organization of Hierarchical Patterns

• Analysis of the time course of the development of the percept in adult observers.

• This analysis is important to understand the processes underlying organization, rather than just the final product of these processes.

Primed matching (Beller, 1971):

• A prime is followed by a pair test of figures.

• Task: Same-different judgment about the test figures.

• The time to respond correctly to same pairs is a function of prime-test similarity.

• Enables to assess implicitly the observer’s perceptual representations.

• If we vary the duration of the prime and construct test figures that are similar to different aspects of the prime, it enables to probe changes in the representation over time.

• Prime-test similarity: Element, configuration, control.

• Prime duration: 40, 90, 130, 390, or 690 ms.

• Comparing responses to test pairs at different prime durations reveals which structures are available in earlier and later representations.

Prime Test Pairs

ElementSimilarity

ConfigurationSimilarity

ElementSimilarity

ConfigurationSimilarity

DifferentSame

Few

-ele

men

tM

any-

elem

ent

Primed Matching

Sequence of Events in a Trial

-70

-60

-50

-40

-30

-20

-10

0

10

20

30

40

40 90 190 390 690

Prime Duration

Pri

min

g (

ms)

Few-Element Many-Element Many-Element:• Configuration is primed

at brief exposures.• Elements are primed at

longer exposures

Few-Element:• Elements are primed at

brief exposures.• Configuration is primed

at longer exposures.

Priming = [RT(ES/Prime)-RT(CS/Prime)] – [RT(ES/Control)-RT(CS/Control]

Priming > 0 Priming of ConfigurationPriming < 0 Priming of Elements

Adapted from Kimchi, 1998

• Task: Search for diamond target among square distractors.

• The target is present at either the global or the local level.

• Display size: 2, 6, or 10 items.

• Main dependent variable: Search rate -the slop of the RT function over display size.

Search time is independent of number of items search is efficient and effortless.

Search time varies with number of items search is inefficient and effortful (involves focused attention).

Visual Search

Many-element:• Global configuration is

searched rapidly and efficiently.

• Local elements are searched slowly and inefficiently.

Few-element:• Local elements are

searched rapidly and efficiently .

• Global configuration is searched less efficiently.

Kimchi, 1998

Results

Target present trials

• Grouping many relatively small elements into a global configuration differs from grouping a few relatively large elements:Grouping many small elements is rapid and

effortlessGrouping a few relatively large elements

consumes time and is effortful.

• Individuation of elements also differs for few versus many elements: Individuation of few large elements is rapid

and effortless Individuation of many small elements -

consumes time and is effortful.

Summary: Microgenesis

• Early and rapid grouping of many small elements and individuation of few large elements – are desirable characteristics for a system

whose one of its goals is object identification and recognition

– because many small elements close to one another are likely to be texture elements of a single object, whereas few large elements are likely to be several discrete objects or several distinctive parts of a complex object.

Notwithstanding the critical role of number and relative size of the elements in the organization of hierarchical patterns, the “nature” of the elements also plays an important role (Han, Humphreys, & Chen, 1999; Kimchi, 1994, 2000)

Interpretation of global advantage: Issues to be considered

• Overall, global advantage is normally observed with the typical hierarchical stimuli (i.e., many-element hierarchical patterns) used in the global–local paradigm to the limits of visibility and visual acuity.

What does it mean? Two issues to be considered:

• Hierarchical patterns provide an elegant control for many intervening variables while keeping the hierarchical structure transparent,

But, the local elements of hierarchical patterns are not the local properties of the global form (Kimchi, 1992, 1994; Navon, 2003).

• Global advantage is not an advantage of a global property of a visual object over its local properties, but rather, an advantage of properties of higher level units over the properties of the lower level units

(Kimchi, 1992). • Global advantage is an advantage of the cluster (or

formation) over its local constituents (Navon, 2003).

• The difference between global and local properties, as operationally defined in the global/local paradigm, may be captured in terms of relative size, and relative size alone may provide a reasonable account for global advantage with hierarchical patterns. – Navon (2003, p. 290) argued that it is a fact of nature

that relative size is “an inherent concomitant of part–whole relationship”.

Yet, if global properties are meant to be properties that depend on the spatial relationship between components, then the difference between global properties and component properties is not captured by their relative size. To distinguish, for example, squareness from its component vertical and horizontal lines, or faceness from its facial components based only on their relative sizes would miss the point.

A refinement of terminology is called for between:

• Global properties defined by the level they occupy within the hierarchical structure of the stimulus

• Holistic/configural properties that arise from the interrelations between the component properties of the stimulus.

The primacy of holistic properties

• Holistic/configural properties: do not inhere in the parts, and cannot be predicted by considering only the individual parts or their simple sum.

Arise on the basis of the interrelations and interactions between the parts (e.g., symmetry, regularity, closure (Garner, 1978; Kimchi, 1992, 1994; Pomerantz, 1981; Rock, 1986; Wagemans, 1995, 1997).

• Exist along with, not instead of, component properties, and are a different aspect of a stimulus (Garner, 1978).

• The hypothesis about the primacy of holistic properties: holistic properties dominate component properties in information processing.

• Examining the relative dominance of component and holistic properties by testing whether the discriminability of the components predicts the discrimination of their configurations.

• If holistic properties dominate information processing, then, irrespective of the discriminability of the components

Discrimination between stimuli that have dissimilar holistic properties should always be easier than discrimination between stimuli that have similar holistic properties

Classification by holistic properties should be easier than classification by the components.

Kimchi,1994

Kimchi,1994

Summary of Results

• The pattern of performance with the configurations was not predicted by the discriminability of the components.

• When both holistic and component properties are present in the stimuli and can be used for the task at hand, performance is dominated by holistic properties, regardless of the discriminability of the component properties.

• When holistic properties are not effective for the task at hand, discrimination and classification can be based on component properties, but there is a significant cost.

No Gap

525550575600625650

40 90 190 390

Res

pons

e T

ime

(ms)

Configuration Similarity

Component Similarity

Configuration Control

Component Control

Small Gap

525550575600625650

40 90 190 390

Res

pons

e T

ime

(ms)

Large Gap

525550575600625650

40 90 190 390

Prime Duration (ms)

Res

pons

e Tim

e (m

s)

Kimchi, 2000

Small Gap

450

500

550

600

650

700

750

40 90 190 390

Prime Duration (ms)

Res

pon

se T

ime

(ms)

Configuration SimilarityComponent SimilarityConfiguration ControlComponent Control

Large Gap

450

500

550

600

650

700

750

40 90 190 390

Prime Duration (ms)

Res

pon

se T

ime

(ms)

Kimchi, 2000

• Studies using primed matching, showed that shapes grouped by closure and/or by collinearity were primed at very short exposure durationsclosure and collinearity were effective

already early in the perceptual processing (Kimchi, 2000; Hadad & Kimchi, 2008).

• Holistic properties were also found to be accessible to rapid search (e.g., Hadad & Kimchi, 2006; Rensink & Enns, 1995).

Kimchi (2000; Hadad & Kimchi, 2008)

Global versus holistic/configural properties

• To examine whether the distinction between global and holistic properties has psychological reality, we must dissociate level of globality (global vs. local) from type of property (configural vs. nonconfigural).

• With hierarchical stimuli, it is possible to construct stimuli in which different types of properties are present at the global and the local levels.

• Relative global or local advantage for many-element hierarchical patterns depends on whether discrimination at each level involves configural or nonconfigural properties.

• When local discrimination involves a configural property like closure, the global advantage markedly decreases or even disappears relative to the case in which discrimination at that level involves a nonconfigural property like orientation (Kimchi,

1994; Han et al., 1999)

Conclusions• Holistic/configural dominance can

arise:– Temporal precedence of the global

level of structure, as when the global configuration of a many-element pattern is represented before the elements are individuated (global precedence)

– Dominance in information processing, as when holistic properties such as closure, dominate component properties in discrimination and classification of visual forms (primacy of holistic properties)

• In light of this evidence, a view that holds that the whole is perceived just by assembling components is hardly tenable.

• Positing holistic dominance as a rigid perceptual law is hardly tenable either– Stimulus factors– Relevance to the task at hand

• Different kinds of wholes with different kinds of parts and part-whole relationships (e.g., a face with its eyes, nose, mouth, and a wall of bricks).

• There are weak or strong wholes, mere aggregation of elements or configuration that preempt the components

• A distinction between global versus local in terms of relative size and levels of representation in a hierarchical structure and between holistic/configural versus component properties.

• Global precedence characterizes the course of processing of some wholes but not of others;

• The processing of some wholes but not of others is dominated by holistic properties

• The processing of some wholes (e.g., faces) is characterized by the integrality of holistic and component properties.

• Holistic dominance and the logical relations between parts and wholes, or between components and configurations: Components are logically prior:

• components can exist without a global configuration, but a configuration cannot exist without components.

• holistic/configural properties do not reside in the component properties but emerge from the interrelations among components.

• The logical structure of the stimulus does not necessarily predict processing consequences at all levels of processing (Garner, 1983; Kimchi, 1992; Kimchi & Palmer, 1985)

• Holistic dominance is also not easily reconciled with the classical view of visual hierarchy in the spirit of Hubel and Wiesel

• The anatomical, structural aspects of the hierarchy of the visual system can be distinguished from the temporal, functional aspects of itnto account the extended connection within cortical areas and the massive feedback pathways (e.g., Maunsell & Essen, 1983).

• Rama Amishav• Orit Baruch• Marlene Behrmann• Tomer Carmel• Aliza Cohen-Savransky• Batsheva Hadad• Steve Palmer• Yossi Pirkner

• Irene Razpurker-Apfeld• Einat Rashal• Suzy Scherf• Guy Sha’ashua• Sarah Shomstein• Branka Spehar• Johan Wagemans• Yaffa Yeshurun

• Max Wertheimer Minerva Center for Cognitive Processes and Human Performance, U. of Haifa

• ISF• BSF

Thank You !!!