Comments on the manuscript: #99-082, S. Lehar: "Harmonic Resonance Theory: An Alternative..."
The manuscript addresses an issue of considerable theoretical significance. On a general level, the reviewer is sympathetic to the view developed in it. The paper presents material which is not universally available and would thus be of interest to a broader public. Also, it is written in a fine stile, though a bit too essayistic relative to its subject. Unfortunately, most of the article is designed as if it were going to offer a solution of the exposed problem while actually it is at best diagnosing the deficiencies of a widely supported view and providing an alternative by way of example, presenting analogies which point out some features of a possible solution. In the argumentation and the sequencing of arguments sufficient rigor is often missing. In some passages of the article the author demonstrates more insight than explicitly stated in assumptions and drawn consequences. There is no discussion of possible fallacies of his own position. In several instances he appears not to be aware of closely related work. Despite of these weaknesses, I would like to encourage the author to submit a revised paper which, while maintaining the general viewpoint, is less ambitious in its goals and, providing a detailed critical review, more analytic and informative.
Major concerns regard the central point of the paper, the concept of Harmonic Resonance, its outline, definition and range of applicability:
(1) While reading the part outlining "Harmonic Resonance Theory" it becomes clear that in fact no theory is presented. What actually is done in some detail from page 8 on is providing a collection of examples of Harmonic Resonance (HR) showing some properties which are analogous to properties of percepts as postulated earlier in the paper. It is a common feature of these examples that the resulting states are Eigen-functions of the given systems under the action of input patterns. Regrettably, not much attention is drawn to specific differences between the chosen examples that would be necessary to pinpoint specificieies of perception more precisely. Indeed, reaction-diffusion systems and observed stages of embryonic development show striking similarities to patterns of micro-genetic development of percepts. While in the case of Chladny figures the resulting states do not depend on the sequencing of inputs, in embryogenesis - as in perception - in passing a sequence of resonant states the system itself develops, and thus the actual sequence of states will result from the interaction of previous states. There is, however, a different aspect in which the behavior of the Chladny system shows stronger similarities to perception than the embryogenetic paradigm does. Namely, it is specific to perception that with a certain precision the states of perceptual systems map outside reality. In the Chladny system this is modeled to some degree by the reaction to additional damping inputs. It should be noted that, if perceptual systems, as suggested, indeed act on the basis of HR, there must be many more specific constraints involved to ensure special "veridicality" properties of the perceptual outcome. Some of them will be specific to modality, and, will, for example, be quite different in audition from those in vision, the modality to which the entire treatment of the paper is restricted. As the paper exclusively proceeds on the basis of examples the difficult analytic problem of concrete modeling of perception are not even touched.
"representation of spatial pattern expressed as patterns of standing waves in a resonating system". This is an interesting subject. However, it is (a) not clarified whether the postulated properties of Gestalts actually follow from this definition or partly derive from additional constraints. Also, there is (b) no serious discussion of possible alternatives which are not based upon standing waves but show similar properties or other properties of potential interest.
ad (a): As for the issue of additional constraints, I doubt that any of the reviewed examples for HR can treat just the case of the dog cited to demonstrate "emergence". For this a hierarchy relation is needed. ON page 19 the author uses such a relation. My point is that this hierarchy relation represents an additional constraint which is not a part of standing wave models per se. In the paper hierarchy formation would have to be introduced earlier as a fundamental property of Gestalts (see remarks below). A similar statement applies to invariance as treated in the paper. The presence of invariance in the Chladny figures is in a way trivial, because it follows as an automatic consequence from the shape of the resonating system. This is an additional material constraint of a type which is untenable in perception where particular symmetries are specific properties of the perceived objects. Therefore, a resonating system like the Chladny plate, in order to show the particular invariance involved, would have to be transformed by the input.
In the introduction attributes like "top-down" (versus bottom-up), "global" (versus "local") or "holistic" (versus element-related) are used connotatively to the approach suggested later. If HR reduces to standing wave modeling this is certainly not correct. The exact place of these properties should be pointed out.
of spatial frequency approaches, in earlier Fourier and later versions like Gabor filters. Although these mechanisms act passively, the global outcome can be equivalent to standing waves. HR closely resembles resonance ideas of Neo-Gibsonian approaches. Neural network approaches like PDP-networks and Grossberg's ART show properties of completion and invariance. In addition, there are proposals to implement truely holistic systems (Pribram) in the sense that any part of them represents a picture in full (though attenuated). The critical difference between HR and these approaches needs to be discussed.
(3) The HR-view is stated as a contrasting view to what the author calls "Neuron Doctrine". For several reasons I share the belief that current receptive field concepts cannot directly account for the emergence of coherent global percepts and that several attempts to do so are seriously misleading. However, notions like the receptive field concept are approximate descriptions of facts. If not accepting receptive fields as a direct basis for explaining perception, we are still left with the problem of why the specific properties of local coding was developed as they did and what their contribution to perception might be. In his remarks the author attributes to local neural organization a very subordinate role. This valuation appears problematic in the view of the apparent universality of some phenomena which have been summarized by principles of receptive field organization. A more likely alternative, not discussed in the article, would seem that organization of perception is dualistic, involving two forms of representations in the brain which are of different, but complementary nature: local excitations as physiologically observable representing the to-be-perceived environment only implicitly, on the one hand, and locally and yet unknown neural equivalents (HR?) of explicit global and coherent percepts, on the other hand. The true puzzle then would be to spell out how these two representations are related in perception as a process. That such a view is possible seems to imply that the contrast between the Neuron Doctrine and its alternative is not necessarily as sharp as suggested in the article.
(4) The paper does not include a discussion of the limitations in the scope of HR. There are clearly such limits: With respect to the reflection of outside reality even in an ideal case HR as described does not provide more than reconstruction of internal analogues of external objects. This does not contribute to the problem of conscious awareness of these objects. To cite Helmholtz, it cannot be seen why "a table in the head" should be more than "a table outside us." As the article at some places is referring to the unity of consciousness, it is not clear what is meant. Another limitation concerns task-dependent recognition: The outlined approach is designed to provide an account of perception in the narrow sense and does not include aspects of task-dependent perceptual processing.
In detail I raise the following concerns:
is presented. This exposition is fairly incomplete, and from the very beginning it puts the issue in a complex context, relating it to Gestalt properties and to the autnor's assumption of "harmonic resonance" long before these notions are actually introduced in the manuscript. In addition, in the same introductory passage, physiological work is prematurely cited to support the author's resonance assumption.
As an outset of theoretical analysis this rather complicated exposition is not very suitable. For this purpose it would seem appropriate (of course, after a brief outline of the main contents of the article) to first properly introduce the Doctrine, its major premises and consequences. In doing so, however, caution is recommended: Firstly, what the author calls the Neuron Doctrine is merely an extreme interpretation of some facts which pose the true puzzle that remains to be solved in order to arrive at a proper theory of perception (see above). Secondly, the doctrine as described by the author is not so absolutely prevailing as he suggests. At least within the realm of the psychology of perception non-local alternatives are widely discussed which are ignored in the article. Early examples are the spatial frequency (or spatio-temporal) approaches based upon Fourier analysis developing since the early seventies, and later versions like Gabor filters. More recent developments include PDP (Parallel Distributed Processing) models of Rumlhard and colleagues, Grossberg's ART network and other neural network approaches.
2. It seems quite natural to show the weaknesses and deficits of the Neuron Doctrine before describing the alternative view. Unfortunately, in the article this is done exclusively on the basis of the argument that the Doctrine cannot account for known Gestalt properties. Although this statement seems true, there exist other, more general statements which are neutral with respect to specific Gestalt properties. For example, this holds true of the above mentioned relation between local neural excitations that represent the environment only implicitly and locally while perception provides an explicit and blobal picture of it. Obviously, the statement that the distributions of local neural excitations cannot directly explain the structure of the related percepts applies to any assumed form of global representation. Consequently, there is no need to build exclusively upon the Gestalt arguments for showing deficiencies of the "Neuron Doctrine". This is important, because it shows that the general problem does not depend on the more specific arguments and thus does not disappear when more specific arguments (as the Gestalt arguments are) would partly turn out to be inadequate. In addition, instead of reducing the problem to a strict contradiction between two views (Neuron Doctrine versus Harmonic Resonance) this way of reasoning would emphasize the necessity of a new interpretation of neural facts within a broader framework. Somehow such an interpretation should also provide a basis for understanding why neural excitations are as they are found.
3. Despite of 2. I agree that fundamental prerequisites of any forthcoming solution can be spellt out starting from an analysis of Gestalts (in a wide sense of the word). Also, emergence, reification, multistability and invariance, as defined by the author, are important points to make.
There are, however, weaknesses to note:
(a) The list of essential properties (as mentioned above) is incomplete, even if viewed under aspects touched later in the paper. Another constitutive property of Gestalts to be included is their potentially hierarchic structure as used on page 19;
(b) important references, in particular to more recent work, are missing whose inclusion would considerably broaden the scope of the paper. I list those that appear important to me:
On p. 6 Koffka is quoted with his physical analogy of the soap bubble. A deeper analysis which is much closer to the intention of the paper was, however, given earlier by Wolfgang Koehler in his 1928 monograph "Physical Gestalts in static and stationary states" (in German) which represents the true scientific origin of field concepts in Gestalt theory. Mathematically, the central concept is that of Eigen functions which also applies to the Chladny figures to which the paper refers.
The aspect of hierarchy has been developed in more recent work of S. Palmer and in the Coding Theory (later Structural Information Theory) of E. Leeuwenberg; the latter approach including also an elegant formulation of minimality in perception referred to in the paper.
The examples of invariance provided in the paper are right, but the indicated range in which the concept applies seems far to narrow. A more complete theory of invariance under discrete transformations, which should be quoted, has been developed in the work of M. Leyton (particularly in "Symmetry, Causality and Mind," 1992). A theory including continuous transformations was suggested earlier by W.C. Hofman.
Overall, it is appropriate to describe properties of percepts which provide relevant constraints for a theory of perceptual processing. However, this description should, firstly, be more complete. Besides a plausible introduction of the viewpoints the paper should, secondly, refer to most advanced knowledge. Thirdly, instead of connecting these descriptions directly with statements about brain processes as is done in the present manuscript I would recommend to present possible conclusions in a separate paragraph. The reason is, as in my criticism above, that this separation would enforce clearness of argumentation and facilitate later revision or specification.
4. Under the heading " Physiological and Psychophysical Evidence (page 13/14) the author cites phenomena, most of which involve periodicities in the time domain, which he takes as evidence of Harmonic Resonance. For the HR proposal it may indeed be important to show that global oscillatory processes exist in the brain which are somehow related to information processing. Whether these oscillations can be shown to fulfill just those functions assumed in the HR approach is, however, a different story. From controversial issues of current discussions caution is recommended. For example, frequencies suggested for binding appear too low to account for the fast micro-genesis of percepts. The hierarchies of oscillations suggested by Geissler and co-workers, on the other hand, seem to serve purposes of temporal co-ordination of the processes involved and there are some indications that they cannot per se account for coding and processing of content.