Similarity and Dissimilarity in Perceptual Organization: On the Complexity of the Gestalt Principle of Similarity

Round 1

Reviewer 1 Report

This is a bit of a visual tone poem celebrating Gestalt grouping. The lead author is a highly respected, extremely creative scientist / artist and the images in the article are very engaging.

 

The main point is well studied across disciplines – discontinuities lead to segmentation – and the presentation of this point as a discovery that dissimilarity is a novel, ignored organizing principle would be a bit jarring to many readers. Dissimilarity is at the opposite end of a continuous scale from similarity, it is hard to claim that it is a different, new thing.

 

Many authors including Grossberg and Mingolla (1985), Hojjatoleslami and Kittler (1998), Pavlidis and Liow (1990), and Julesz (1981) with his texton theory, all looked at how similarity allowed regions to be grouped together up to a discontinuity which then segmented the grouped areas. That discontinuity, or low similarity, or dissimilarity, across a border is the focus of Pinna et al.’s contribution here but it seems that they feel that they are the first to notice this. They need to integrate their observations with this vast literature. It is important to note that the discontinuity (or dissimilarity) that triggers a segmentation is relative. A small shift in colour between two uniformly coloured sets of elements may be enough, but if the elements have some variability, a larger discontinuity is required for segmentation. Dissimilarity that leads to image organization is relative not absolute.

 

The grouping of the isolated, odd coloured dots in Figure 8 is part of the pop-out literature that followed Treisman and Gormican’s findings. It is not a challenge to the similarity principle. In fact, this odd-ball, pop-out literature assumes that it is the grouping of the similar elements that makes the odd items so salient.  Again this needs integration with another large literature.

 

I do not know what to make of the reading section. Letters and word spaces have evolved to be readable. The variations of spacing, light vs dark, and dots may be telling us something but in the reading literature, this would be accompanied by data. These informal observations, none of which appear to be unexpected, would never stand on their own. This section should be removed.

 

The dotted squares of Figure 35 may be a new finding that would deserve more emphasis. The many other shape distortions that preceded this are intriguing but do not seem to build to a point. This could use better organization.

 

This article is a palimpsest of delightful images and conjectures with a serious lack of contact with other relevant literatures.

Author Response

REVIEWER #1

This is a bit of a visual tone poem celebrating Gestalt grouping. The lead author is a highly respected, extremely creative scientist / artist and the images in the article are very engaging.The main point is well studied across disciplines – discontinuities lead to segmentation – and the presentation of this point as a discovery that dissimilarity is a novel, ignored organizing principle would be a bit jarring to many readers. Dissimilarity is at the opposite end of a continuous scale from similarity, it is hard to claim that it is a different, new thing.
We thank Reviewer’s comments, giving us the opportunity to define more accurately and formally the notion of dissimilarity and the differences with similarity. Within the ms, we have now extended the discussion as follow:
“Our work showed that the use of similarity alone is not sufficient to fully and correctly understand the dynamics of grouping and, more generally, of perceptual organization occurring both in the classical and, particularly, in the new phenomena here presented. Limits and incompleteness of the similarity principle revealed the basic, more general and stronger role of dissimilarity in perceptual grouping and figure-ground organization under a large variety of conditions.
The results suggest that both similarity and dissimilarity can deliver and solve real and useful issues in perceptual organization. The two attributes are, not only the opposite ends of a continuum, but primarily distinct factors of formation object. Dissimilarity was shown as a basic principle of figure-ground segregation, as a tool able to create at will new groups and visual objects within patterns where they are totally invisible, as an attribute that is able to accentuate different shape components within the same object, as a way to distort shapes, to create visual illusions, but also to reduce or annul them and, finally, to decompose, ungroup and reshape single objects. The results demonstrated the necessity to introduce a principle of dissimilarity and complement the similarity already studied by Gestalt psychologists.
More in details, dissimilarities have been shown to be responsible for creating segregation, distinction and separation between adjacent groups, similarities for putting homogeneously together the elements within each group. Dissimilarities are aimed to divide boundaries and to emphasize discontinuities among regions of the visual field, similarities to enhance the inner homogeneity. In short, dissimilarities highlight and segregate the boundaries of an object, similarities reveal its inner surface qualities, like the grain and the matter.
Metaphorically, we can think of two groups of people, like, for example, supporters to two opposing teams. The dynamics between dissimilarity and similarity highlight and magnify differences between the two groups of supporters and reduce or minimize differences among the members within each group. Analyzing the outcomes from the two different perspectives, dissimilarities are perceived in between the teams, eliciting a strong segregation and differentiation effect, similarities are seen within the members of each group inducing unification and uniformity.
The dissimilarity/similarity dynamics are very effective to explain complex human interactions like, for instance, racism, extremism, political, sexual or age conflicts, etc. Moreover, a small dissimilarity in one state can result in large differences in a later state (see Figs. 35-39). A deep understanding of these dynamics cannot be based only on the similarity. Similarity is not sufficient. In short, dissimilarities create boundaries, barriers, differences, divisions, reactions to the changes and surprise, while similarities produce uniformness, homogeneity and unity.
There is a further and more phenomenological demonstration of the distinction between the two factors. This is the spontaneous and independent use of the two terms in our language to describe perceptual organization both from the point of view of what emerges as a distinct informative separation and differentiation and from the point of view of what appears as a uninformative uniformity and homogeneity.
In summary, we can phenomenally define dissimilarity as the information content related to all kinds of changes, breaks and discontinuities occurring along and within a potential homogeneity and continuity. In this sense, dissimilarity is a gradient boosting, an abrupt break of uniformity producing “surprise”. More formally, dissimilarity can be associated with an unexpected discontinuity function occurring in the derivative of a gradient of visual attributes (cfr. Mach bands – Ratliff, 1965; Kingdom, 2014 – and the watercolor illusion – Pinna, 2008).
If an abrupt change is unlikely, with a low probability of occurring, and produce “surprise”, then we gained more information from this event than if the event had been something we were expecting. In essence, an unexpected event produce more information by changing our perception of the world. Dissimilarities can be considered as the information content of an event and as such it can be phenomenally measured.
There is a more general and intriguing point that can become a challenging source of discussion at this stage. It is related to the connections between dissimilarity, the notion of visual information and information in the sense of the theory of information.
In short, Shannon (1948) quantified the amount of information in a signal, suggesting that it is the amount of unexpected data contained in the message. He considered information in that which is not random; therefore information is not noise, but what unpredictably adds information. In a given set of possible data, the information of a message describing one of these data quantifies the symbols needed to encode the data in an optimal way. The term ‘optimal’ means that the obtained code word will determine the datum unambiguously, isolating it from all others in the set, and will have minimal length, namely a minimal number of symbols. Information theory is based on the measure of uncertainty in term of Shannon entropy (H) that is similar to the definition of entropy in thermodynamics according to Boltzmann’s principle. In simple terms, the entropy can be described as how long of a message (in bits) we need to convey the value of the stimulus.
Also interesting for our purposes is the way Bateson (1979) defined “information”, incapsulated in his famous motto: information is a difference that makes a difference.
The notion of dissimilarity is likely the phenomenal attribute closest to the notion of information as previously described. Given outcomes within a probability space and the information content inherent in those outcomes, entropy denote the expected information over the outcomes, i.e., the average information over all of the possible outcomes. Since dissimilarity is the information that describes the degree of surprise of an event, the entropy measures, on average, how surprised we are going to be by the outcomes.
Dissimilarities that make dissimilarities, paraphrasing Bateson, create and deliver objects and information reducing uncertainty and entropy. Therefore, according to information theory, entropy is in our conditions a measure of the degree of uniformness of a variation within a pattern of stimuli. The higher the entropy, the closer the visual pattern is to having all of its outcomes being equally likely.
The attribute of dissimilarity, developed in this work, although may appear close to the notion of information, it does not stand for it, even if it might be a good candidate. Deeper and more focussed demonstrations are required together with more crucial stimuli to go closer and closer to this conclusion. At the moment this is just an impromptu hint, a first look to be further explored.”

Within the ms, there are many examples demonstrating that similarity is insufficient to explain grouping and perceptual organization under our conditions (see. Figs. 28, 29, 31, 32, 34, 35, 39). Here, the accentuation depends mostly or uniquely on dissimilarity. The more dissimilar is the accent the stronger is the resulting accentuation.

Many authors including Grossberg and Mingolla (1985), Hojjatoleslami and Kittler (1998), Pavlidis and Liow (1990), and Julesz (1981) with his texton theory, all looked at how similarity allowed regions to be grouped together up to a discontinuity which then segmented the grouped areas. That discontinuity, or low similarity, or dissimilarity, across a border is the focus of Pinna et al.’s contribution here but it seems that they feel that they are the first to notice this. They need to integrate their observations with this vast literature. It is important to note that the discontinuity (or dissimilarity) that triggers a segmentation is relative. A small shift in colour between two uniformly coloured sets of elements may be enough, but if the elements have some variability, a larger discontinuity is required for segmentation. Dissimilarity that leads to image organization is relative not absolute.
We have incorporated the references suggested by the Reviewer. We agree that dissimilarity is relative not absolute, however we are convinced (see the previous point) that a distinction between similarity and dissimilarity is supported by data and useful to understand complex phenomena like those illustrated in the ms.

The grouping of the isolated, odd coloured dots in Figure 8 is part of the pop-out literature that followed Treisman and Gormican’s findings. It is not a challenge to the similarity principle. In fact, this odd-ball, pop-out literature assumes that it is the grouping of the similar elements that makes the odd items so salient.  Again this needs integration with another large literature.
In Fig. 8 the rational is exactly the opposite: it is not the grouping of the similar elements that makes the odd items so salient, but rather the odd items define or accentuate the grouping of all the black dots.

I do not know what to make of the reading section. Letters and word spaces have evolved to be readable. The variations of spacing, light vs dark, and dots may be telling us something but in the reading literature, this would be accompanied by data. These informal observations, none of which appear to be unexpected, would never stand on their own. This section should be removed.
The rationale of the reading section is essential to clarify that the directional and sequential nature of reading from left to right is useful to unambiguously demonstrate the peculiar role of dissimilarity as a principle different from similarity. In fact, while reading what one meets first is a dissimilar letter. Therefore, the first attribute is the dissimilarity, not the similarity. Again the dissimilarity first of all segregates, while the similarity put together, just after the segregation. We think that the reading section is essential, since it is the most effective condition and the unique demonstration of sequential grouping. For this reason we suggest to keep it.

The dotted squares of Figure 35 may be a new finding that would deserve more emphasis. The many other shape distortions that preceded this are intriguing but do not seem to build to a point. This could use better organization.
This condition produces also apparent motion, now included within the text. Moreover, this further effect demonstrating that dissimilarity cannot be considered only as the opposite pole of similarity (see the full answer to the first point).

 

Reviewer 2 Report

I do not find any major issues in this paper. I only have minor comments and questions.

1) It would be nice to define a bit more what is the defining core property of dissimilarity a bit more formally. I state this because it is not obvious that a same principle is guiding the many example provided. In the examples, there are many different physically measurable features that vary: color, polarity, contrast, etc. What is the unitary "rule" that subsumes all these examples under the same phenomenological effect, irrespective of the different physical features that vary?

2) The interesting discussion on information and entropy I think could be more complete if other notions such as "constraint" or "bias" in perceptual interpretation could be discussed. These are also equally good candidates.

3) I wonder if dissimilarity cannot be simply defined as a change or break of potential organization. In many of the examples provided, I can go forth and back like in a bi-stable perceptual process or interpretation of the alternative configurations, and the effect of the dissimilarity is contained itself in a larger "meta-frame" provided by the most obvious reading of the groupings as forming an underlying complete whole. It seems dissimilarity points at an incompleteness in the organization that can only be overcome with another interpretation of completeness. 

4) In which way is dissimilarity different from changing the way the observer distributes selective attention to familiar patterns of organization?

Author Response

REVIEWER #2

I do not find any major issues in this paper. I only have minor comments and questions. 1) It would be nice to define a bit more what is the defining core property of dissimilarity a bit more formally. I state this because it is not obvious that a same principle is guiding the many example provided. In the examples, there are many different physically measurable features that vary: color, polarity, contrast, etc. What is the unitary "rule" that subsumes all these examples under the same phenomenological effect, irrespective of the different physical features that vary?
We thank Reviewer’s comments and suggestions. Within the ms, we have now extended the Discussion section as follow:
“Our work showed that the use of similarity alone is not sufficient to fully and correctly understand the dynamics of grouping and, more generally, of perceptual organization occurring both in the classical and, particularly, in the new phenomena here presented. Limits and incompleteness of the similarity principle revealed the basic, more general and stronger role of dissimilarity in perceptual grouping and figure-ground organization under a large variety of conditions.
The results suggest that both similarity and dissimilarity can deliver and solve real and useful issues in perceptual organization. The two attributes are, not only the opposite ends of a continuum, but primarily distinct factors of formation object. Dissimilarity was shown as a basic principle of figure-ground segregation, as a tool able to create at will new groups and visual objects within patterns where they are totally invisible, as an attribute that is able to accentuate different shape components within the same object, as a way to distort shapes, to create visual illusions, but also to reduce or annul them and, finally, to decompose, ungroup and reshape single objects. The results demonstrated the necessity to introduce a principle of dissimilarity and complement the similarity already studied by Gestalt psychologists.
More in details, dissimilarities have been shown to be responsible for creating segregation, distinction and separation between adjacent groups, similarities for putting homogeneously together the elements within each group. Dissimilarities are aimed to divide boundaries and to emphasize discontinuities among regions of the visual field, similarities to enhance the inner homogeneity. In short, dissimilarities highlight and segregate the boundaries of an object, similarities reveal its inner surface qualities, like the grain and the matter.
Metaphorically, we can think of two groups of people, like, for example, supporters to two opposing teams. The dynamics between dissimilarity and similarity highlight and magnify differences between the two groups of supporters and reduce or minimize differences among the members within each group. Analyzing the outcomes from the two different perspectives, dissimilarities are perceived in between the teams, eliciting a strong segregation and differentiation effect, similarities are seen within the members of each group inducing unification and uniformity.
The dissimilarity/similarity dynamics are very effective to explain complex human interactions like, for instance, racism, extremism, political, sexual or age conflicts, etc. Moreover, a small dissimilarity in one state can result in large differences in a later state (see Figs. 35-39). A deep understanding of these dynamics cannot be based only on the similarity. Similarity is not sufficient. In short, dissimilarities create boundaries, barriers, differences, divisions, reactions to the changes and surprise, while similarities produce uniformness, homogeneity and unity.
There is a further and more phenomenological demonstration of the distinction between the two factors. This is the spontaneous and independent use of the two terms in our language to describe perceptual organization both from the point of view of what emerges as a distinct informative separation and differentiation and from the point of view of what appears as a uninformative uniformity and homogeneity.
In summary, we can phenomenally define dissimilarity as the information content related to all kinds of changes, breaks and discontinuities occurring along and within a potential homogeneity and continuity. In this sense, dissimilarity is a gradient boosting, an abrupt break of uniformity producing “surprise”. More formally, dissimilarity can be associated with an unexpected discontinuity function occurring in the derivative of a gradient of visual attributes (cfr. Mach bands – Ratliff, 1965; Kingdom, 2014 – and the watercolor illusion – Pinna, 2008).
If an abrupt change is unlikely, with a low probability of occurring, and produce “surprise”, then we gained more information from this event than if the event had been something we were expecting. In essence, an unexpected event produce more information by changing our perception of the world. Dissimilarities can be considered as the information content of an event and as such it can be phenomenally measured.
There is a more general and intriguing point that can become a challenging source of discussion at this stage. It is related to the connections between dissimilarity, the notion of visual information and information in the sense of the theory of information.
In short, Shannon (1948) quantified the amount of information in a signal, suggesting that it is the amount of unexpected data contained in the message. He considered information in that which is not random; therefore information is not noise, but what unpredictably adds information. In a given set of possible data, the information of a message describing one of these data quantifies the symbols needed to encode the data in an optimal way. The term ‘optimal’ means that the obtained code word will determine the datum unambiguously, isolating it from all others in the set, and will have minimal length, namely a minimal number of symbols. Information theory is based on the measure of uncertainty in term of Shannon entropy (H) that is similar to the definition of entropy in thermodynamics according to Boltzmann’s principle. In simple terms, the entropy can be described as how long of a message (in bits) we need to convey the value of the stimulus.
Also interesting for our purposes is the way Bateson (1979) defined “information”, incapsulated in his famous motto: information is a difference that makes a difference.
The notion of dissimilarity is likely the phenomenal attribute closest to the notion of information as previously described. Given outcomes within a probability space and the information content inherent in those outcomes, entropy denote the expected information over the outcomes, i.e., the average information over all of the possible outcomes. Since dissimilarity is the information that describes the degree of surprise of an event, the entropy measures, on average, how surprised we are going to be by the outcomes.
Dissimilarities that make dissimilarities, paraphrasing Bateson, create and deliver objects and information reducing uncertainty and entropy. Therefore, according to information theory, entropy is in our conditions a measure of the degree of uniformness of a variation within a pattern of stimuli. The higher the entropy, the closer the visual pattern is to having all of its outcomes being equally likely.
The attribute of dissimilarity, developed in this work, although may appear close to the notion of information, it does not stand for it, even if it might be a good candidate. Deeper and more focussed demonstrations are required together with more crucial stimuli to go closer and closer to this conclusion. At the moment this is just an impromptu hint, a first look to be further explored.”

2) The interesting discussion on information and entropy I think could be more complete if other notions such as "constraint" or "bias" in perceptual interpretation could be discussed. These are also equally good candidates.
Please see the extension of the Discussion section.

3) I wonder if dissimilarity cannot be simply defined as a change or break of potential organization. In many of the examples provided, I can go forth and back like in a bi-stable perceptual process or interpretation of the alternative configurations, and the effect of the dissimilarity is contained itself in a larger "meta-frame" provided by the most obvious reading of the groupings as forming an underlying complete whole. It seems dissimilarity points at an incompleteness in the organization that can only be overcome with another interpretation of completeness.
Right. We agree on reviewer suggestions now incorporated within the ms (see the first point).

4) In which way is dissimilarity different from changing the way the observer distributes selective attention to familiar patterns of organization?
On the basis of our data we cannot answer to this interesting questions, that deserves further studies. Thanks for this suggestion.

Reviewer 3 Report

Overall this is an informative and important manuscript. Gestalt organization principles have been around for a century, yet their use in contemporary research has been limited despite the fact that these principles are central to our understanding of visual processing. Part of this problem is that it has not been possible to express these principles in clear (preferably mathematical) language. This paper focuses on the principle of similarity and shows how this seemingly simple concept runs into problems when examined closely. Some may view dissimilarity just as one side of a continuous measure that includes both similarity and dissimilarity; however, the examples provided in the manuscript show that similarity and dissimilarity may control different processes and hence need to be considered and analyzed accordingly.

One recommendation I have for the authors would be to link these observations to research in other areas so that these important concepts get adopted widely by researchers outside of the Gestalt community. For example, the discussion of "pop up" may be linked to the large literature on attention and search. The discussion of perceived shape (square vs diamond) can be linked to reference-frames and object recognition. Examples with text are linked to reading; but connecting it to this literature directly with relevant reading-related references would be helpful. Similarly, spirals and depth perception: add references to the relevant literature on depth perception.The idea about information theory may cite vision approaches that use information-theoretic concepts (e.g., sparse coding).

To sum, this is a very good paper on Gestalt psychology, but I believe it will have more impact if the concepts are related to other research areas that can benefit from wider use of these concepts.

In general the manuscript is well written but it may need a re-reading:

Examples: p. 2 the sentence that starts with Easily - the subject is missing, add "It". P. 4. 2nd parag., 3rd line: "living" should read "leaving. P. 11, parag. 2, line 2: pattern should read patterns. P. 17, line 6 from the bottom: results should read result.

Author Response

REVIEWER #3

Overall this is an informative and important manuscript. Gestalt organization principles have been around for a century, yet their use in contemporary research has been limited despite the fact that these principles are central to our understanding of visual processing. Part of this problem is that it has not been possible to express these principles in clear (preferably mathematical) language. This paper focuses on the principle of similarity and shows how this seemingly simple concept runs into problems when examined closely. Some may view dissimilarity just as one side of a continuous measure that includes both similarity and dissimilarity; however, the examples provided in the manuscript show that similarity and dissimilarity may control different processes and hence need to be considered and analyzed accordingly.
One recommendation I have for the authors would be to link these observations to research in other areas so that these important concepts get adopted widely by researchers outside of the Gestalt community. For example, the discussion of "pop up" may be linked to the large literature on attention and search. The discussion of perceived shape (square vs diamond) can be linked to reference-frames and object recognition. Examples with text are linked to reading; but connecting it to this literature directly with relevant reading-related references would be helpful. Similarly, spirals and depth perception: add references to the relevant literature on depth perception.The idea about information theory may cite vision approaches that use information-theoretic concepts (e.g., sparse coding).
To sum, this is a very good paper on Gestalt psychology, but I believe it will have more impact if the concepts are related to other research areas that can benefit from wider use of these concepts.
According to the Reviewer’s suggestions have now extended the references. Thanks a lot.

In general the manuscript is well written but it may need a re-reading:
Examples: p. 2 the sentence that starts with Easily - the subject is missing, add "It". P. 4. 2nd parag., 3rd line: "living" should read "leaving. P. 11, parag. 2, line 2: pattern should read patterns. P. 17, line 6 from the bottom: results should read result.
We fixed these and more typos as suggested.

Round 2

Reviewer 1 Report

This is now fine.